INTRODUCTION INTO QCD SUM RULE APPROACH

A.V. RADYUSHKIN

Physics Department, Old Dominion University, Norfolk, VA 23529and

Theory Group, Jefferson Lab, Newport News, VA 23606

In these lectures, I describe basic techniques of the QCD sum rule approach. Thebasic concepts of the approach are introduced using a simple model of quantum-mechanical oscillator in 2+1 dimensions. Then I discuss their field-theoreticalextension and the construction of the operator product expansion for current cor-relators in QCD. The calculation of static parameters is illustrated on the exampleof sum rules in the vector and axial meson channels. Finally, the QCD sum rulecalculation of the pion electromagnetic form factor is presented as an example ofapplication of the method to dynamic characteristics of the hadrons.

1 Introduction

1.1 Introductory remarks.

QCD sum rules, invented more than 20 years ago by Shifman, Vainshteinand Zakharov1 are still a rather popular approach for theoretical study of thehadronic structure. Applied originally to the simplest static hadronic char-acteristics, like masses, leptonic widths etc., they were also used to calculatemuch more complicated things like hadronic wave functions and form factors.

To begin with, I will briefly discuss the basic questions• Why we need QCD sum rules?• What are QCD sum rules?• How the QCD sum rules are working?using for illustration the QCD calculation of the pion form factor. In subse-quent sections, I will discuss in more detail the machinery of QCD sum rules inseveral cases: calculation of the properties of the lowest resonances in ρ-mesonand π/A1 channels, and analysis of the pion form factor in various kinematicalsituations.

1.2 Pion form factor in QCD.

According to the factorization theorem2,3, for sufficiently large momentumtransfers, one can represent the pion form factor as a sum of terms of increasingcomplexity (see Fig.1). The first term (purely soft contribution, Fig.1a con-tains no short-distance (SD) subprocesses. The second term Fig.1b contains a

1

d)c)

b)a)

+++

+=

Figure 1: Structure of factorization for the pion form factor in QCD

hard gluon exchange. There are also corrections to the hard term: higher or-der corrections Fig.1c containing extra αs factors and higher twist correctionsFig.1d. In perturbative QCD one can calculate only the hard terms of thisexpansion.

The first result of perturbative QCD for exclusive processes is the pre-diction for the asymptotic behavior of the pion electromagnetic form factor4,2,5,6:

Fπ(Q2) =∫ 1

0

dx

∫ 1

0

t(xP, (1− x)P ; yP ′, (1− y)P ′, q) ϕ(x) ϕ(y) dy , (1)

where ϕ(x) is the pion wave function giving the probability amplitude to findthe pion in a state where quarks carry fractions xP and (1 − x)P of its lon-gitudinal momentum P ; q = P ′ − P is the momentum transfer to the pionand T (. . .) is the perturbatively calculable short-distance amplitude. The wavefunction ϕ(x) accumulates long-distance information about the pion structureand cannot be calculated within the pQCD approach. Normally, some phe-nomenological assumptions are used about the form of this function dependingon a single parameter x. Still, there is a challenge to calculate it from firstprinciples of QCD.

To calculate the soft contribution to the form factor one should knowsomething like the light-cone wave function ψ(x, k⊥) depending also on thetransverse momentum k⊥. I say “something like”, because in principle it isimpossible to prove that the soft term reduces in QCD to the convolution

F (q2) ∼∫ψP (x, k⊥)ψ∗P (x, k⊥ + xq) d2k⊥dx. (2)

2

In fact, one should treat the soft term as a nonfactorizable amplitude and theproblem is to calculate it in some reliable way.

The QCD sum rules are just giving the method to calculate the nonpertur-bative hadronic characteristics incorporating the asymptotic freedom propertyof QCD in a way similar to pQCD.

1.3 Basic ideas of the QCD sum rule approach

Among the existing approaches to the analysis of the nonperturbative effectsin QCD the most close to perturbative QCD is the QCD sum rule method1. Let us formulate its basic ideas within the context of the pion form factorproblem.

It is evident that one cannot directly study the soft contribution with theon-shell pions, because then only long distances are involved. But perturbativeQCD can be applied in a situation when all relevant momenta q, p1, p2 arespacelike and sufficiently large: |q2|, |p2

1|, |p22| > 1 GeV2. To describe the virtual

pions one should use some interpolating field, the usual (for the QCD sum rulepractitioners) choice being the axial current jα5 = dγ5γ

αu. Its projection ontothe pion state |P, π〉 is proportional to fπ :

〈0|jα5 |P, π〉 = ifπPα. (3)

Via the dispersion relation

T (p21, p

22, q

2) =1π2

∫ ∞0

ds1

∫ ∞0

ds2ρ(s1, s2, q

2)(s1 − p2

1)(s2 − p22)

(4)

one can relate the amplitude T (p21, p

22, q

2) to its time-like counterpart ρ(s1, s2, q2)

containing the double pole term

ρπ(s1, s2, q2) = π2f2

πδ(s1 −m2π)δ(s2 −m2

π)Fπ(Q2) (5)

corresponding to the pion form factor. However, the axial current has nonzeroprojections onto other hadronic states (A1−meson, say) as well, and the spec-tral density ρ(s1, s2, q

2) contains also the part ρhigher states(s1, s2, q2) related

to other elastic and transition form factors. This is the price for going off thepion mass shell. The problem now is to pick out the Fπ term from the wholemess.

Of course, calculating T (p21, p

22, q

2) in the lowest orders of perturbationtheory one never observes something like the pion pole: one obtains a smoothfunction ρpert(s1, s2, q

2) corresponding to transitions between the free-quark

3

+++

b) c)

=

a)

Figure 2: Structure of the operator product expansion for T (p21, p

22, q

2)

ud−states with invariant masses s1 and s2, respectively. The difference be-tween “exact” density ρ(s1, s2, q

2) and its perturbative analog ρpert(s1, s2, q2)

is reflected by additional nonperturbative contributions to T (p21, p

22, q

2). Thesecontributions are due to quark and gluon condensates 〈qq〉, 〈GG〉 etc., de-scribing (and/or parameterizing) the nontrivial structure of the QCD vacuumstate. Formally, these terms appear from the operator product expansion forthe amplitude T (p2

1, p22, q

2) (Fig.2):

T (p21, p

22, q

2) = T pert(p21, p

22, q

2) + a〈GG〉(p2)3

+ bαs〈qq〉2(p2)4

+ . . . . (6)

The problem now is to construct such a model of the spectral densityρ(s1, s2, q

2) which gives the best agreement between two expressions for T , i.e.,Eqs. (4) and (6). Naturally, having only a few first terms of the 1/p2−expansionone can hope to reproduce only the gross features of the hadronic spectrum inthe relevant channel. Still, just using the simple fact that the condensate con-tributions die out for large p2, one obtains the global duality relation betweenquark and hadronic densities∫ ∞

0

ds1

∫ ∞0

ds2

(ρ(s1, s2, q

2)− ρpert(s1, s2, q2))

= 0. (7)

In other words, integrally the two densities are rather similar. One can evenexpect that they are very close for high s to secure the convergence of theintegral (7).

The standard ansatz is to approximate the higher states contribution intoρ(s1, s2, q

2) by the free quark density (compare to the quasiclassical approxi-mation hor high levels in quantum mechanics):

ρhigher states(s1, s2, q2) = [1− θ(s1 < s0)θ(s2 < s0)] ρpert(s1, s2, q

2), (8)

with s0 being the effective threshold for the higher states production.

4

1.4 Summary

To summarize, the basic ingredients of the QCD sum rule approach are• Operator product expansion, or factorization of the short-distance, perturba-tively calculable amplitudes, with the long-distance information accumulatedby the quark and gluon condensates – the basic parameters of the approach –characterizing the properties of the QCD vacuum.• Dispersion relations, relating the perturbatively calculated amplitude withthe physical hadronic parameters.• To extract the hadronic parameters one should construct an ansatz for thehadronic spectrum (like “first resonance plus continuum”, etc.) and thenperform fitting procedure.

Just from the above list it is clear that QCD sum rules cannot produceexact results. Their precision is unavoidably limited by the following facts :• One can calculate normally only few first terms of the OPE.• The model ansatz reproduces the real density only approximately.• Furthermore, in some channels could appear nonperturbative short-distancecontributions (“direct instantons”).

As a result, using the QCD sum rules, one can normally expect only 10%- 20% accuracy.

2 QCD Sum Rules in Quantum Mechanics

We are going to apply the QCD sum rule method in the theory, where noexact results for hadrons were obtained so far. But, as formulated above, theQCD sum rule method is quite general. In fact, many features of the methodcan be demonstrated on the exactly solvable model of the quantum-mechanicaloscillator. In the discussion below, we basically follow the pioneering paper7,with some simplifying modifications8,9.

2.1 Sum rules for the lowest state of two-dimensional oscillator.

Let us consider the simplest example of a system with confinement, the oscil-lator, with the confining potential being

V (~x) =mω2

2r2. (9)

Formulas are the simplest (without square roots, complicated factorial terms,etc.) if we choose the oscillator in two spatial dimensions (+ time dimension).In this case the energy levels are given by

En = (2n+ 1)ω, (10)

5

and the values of the wave functions at the origin (quantum-mechanical ana-logues of fπ and other leptonic decay parameters) are the same for all levelscorresponding to radial excitations with zero angular momentum:

|ψn(0)|2 =mω

π. (11)

To calculate E0 and ψ0(0) using the sum rule approach one should proceedin a rather peculiar way: instead of concentrating on the lowest level, oneshould consider the weighted sum of all levels:

M(ε) =∞∑k=0

|ψk(0)|2e−Ek/ε (12)

in which the lowest level dominates only for very small values of ε. For large εall levels are essential. This combination (12) is directly related to the two-timeGreen function

G(x1, t1|x2, t2) =∞∑k=0

ψ∗k(x2)ψk(x1)e−iEk(t2−t1) (13)

describing the probability amplitude for transition from the point x1 at timet1 to the point x2 at time t2. To get M(ε) one should take x2 = x1 = 0, t1 = 0and t2 = 1/iε. In other words, the transition is taken in the imaginary time.

In our case, explicit form of M(ε) can be easily calculated using Eqs. (10)and (11):

Mosc(ε) =∞∑k=0

(mωπ

)e−(2k+1)ω/ε =

mω

π sinh(ω/ε)(14)

In the large-ε limit one has

Mosc(ε) =mε

2π

(1− ω2

6ε2+

7360

ω4

ε4− 31

15120ω6

ε6+ . . .

)(15)

i.e., the first term of the large-ε (≡small-time) expansion has no dependenceon ω, the paramater characterizing the strength of oscillator potential. Thissimply means that at small time differences the Green function coincides withthe free one:

G(2|1)free =m

2πi(t2 − t1)exp

−im(x2 − x1)2

2(t2 − t1)

(16)

andMfree(ε) =

mε

2π(17)

6

The corrections to the free-particle behavior in Eq. (15) are in powers of ω2.This corresponds to the perturbation expansion in the potential V (r) (9):

M = M (free)+M (free)⊗V ⊗M (free)+M (free)⊗V ⊗M (free)⊗V ⊗M (free)+. . .(18)

The function Mfree(ε) can also be represented as a sum over the (free)states with Ek = k2/2m:

Mfree(ε) =∫

d2k

(2π)2e−Ek/ε =

m

2π

∫ ∞0

e−E/εdE ≡ 1π

∫ ∞0

e−E/ερ(E)dE .

(19)We introduced here the new function

ρfree(E) =m

2θ(E > 0), (20)

the spectral density corresponding to the free motion, which will play an im-portant role in the subsequent discussion. The oscillator function M(ε) alsocan be written in the spectral representation with the spectral density ρ(E)which is in this case a superposition of the delta-functions:

ρosc(E) = mω∞∑k=0

δ(E − (2k + 1)ω). (21)

At first sight, the two densities are completely different. However, if one in-tegrates them from zero to the midpoint between the first and second levels,one obtains the same result:

1m

∫ 2ω

0

ρfree(E)dE = ω =1m

∫ 2ω

0

ρosc(E)dE . (22)

Moreover, similar relations hold for other levels:∫ 2(k+1)ω

2kω

ρfree(E)dE = mω =∫ 2(k+1)ω

2kω

ρosc(E)dE . (23)

The equality is not violated even if one multiplies ρ(E) by an extra power ofthe energy E:∫ 2(k+1)ω

2kω

ρfree(E)EdE = mω2(2k + 1) =∫ 2(k+1)ω

2kω

ρosc(E)EdE . (24)

Thus, there exists the local duality between each resonance and the free states.It looks like each level “takes” its part of the free spectral density, leaving theintegral unchanged.

7

Turning back to the ω-expansion for Mosc(ε) (Eq. (15))

Mosc(ε) = Mfree(ε)

1− ω2

6ε2+

7360

ω4

ε4− 31

15120ω6

ε6+ . . .

(25)

one can write

Mosc(ε)−Mfree(ε) = O(ω2/ε) +O(ω4/ε3) + . . . . (26)

or, in terms of the spectral densities

1π

∫ ∞0

[ρosc(E) − ρfree(E)

]e−E/εdE =

∑k=1

Akεk. (27)

The terms on the rhs, suppressed for large ε by powers of 1/ε are normallyreferred to as “power corrections”. Thus, the difference between the integratedspectral densities vanishes as ε→∞:

1π

∫ ∞0

[ρosc(E)− ρfree(E)]dE = 0 , (28)

since there is no O(1) term in the rhs of Eq. (27). Eq. (28) is a standard globalduality relation. In quantum mechanics one can obtain such a relation for anypotential regular at the origin.

The lesson is that the two densities, despite the apparent absence of sim-ilarity, are rather close to each other in an integral sense. Now, the idea is toincorporate this similarity in the sum rule

|ψ0(0)|2e−E0/ε+“higher states” =mε

2π

1− ω2

6ε2+

7360

ω4

ε4− 31

15120ω6

ε6+ . . .

(29)

by substituting the “higher states” contribution by that of the free states hav-ing energy above some threshold s0. The sum rule then reads

|ψ0(0)|2e−E0/ε =ε

2

(1− e−s0/ε

)− ω2

12ε+

7720

ω4

ε3− 31

30240ω6

ε5+ . . . (30)

where|ψ0(0)|2 =

π

m|ψ0(0)|2

The parameters to be extracted from the sum rule are E0, ψ0(0) and s0. Thefirst step is to extract E0. To this end we differentiate Eq. (30) with respectto (−1/ε) to get

|ψ0(0)|2E0e−E0/ε =

ε2

2− ε

2

2

(1 +

s0

ε

)e−s0/ε+

ω2

12− 7

240ω4

ε2+

316048

ω6

ε4+. . . (31)

8

The ratio of the two equations just gives the expression for E0:

E0 =ε2

2 −ε2

2 (1 + s0ε )e−s0/ε + ω2

12 −7

240ω4

ε2 + 316048

ω6

ε4 + . . .ε2 (1− e−s0/ε)− ω2

12ε + 7720

ω4

ε3 −31

30240ω6

ε5 + . . .. (32)

If we would take the whole series in ω/ε and the exact expression for thehigher states contribution, then the result for E0 would be the ε-independentexact value E0 = ω. However, under approximations we made, the E0-valueextracted from Eq. (32) depends both on the auxiliary parameter ε and thethreshold parameter s0. The standard strategy is to take the value of s0,for which E0 has the minimal dependence on ε in the intermediate region1 < ε/ω < 2 where both the higher power corrections and the error due tothe rough model of the higher states contribution are small (≈ 5 − 10%).Truncating the series by terms of the second order in ω2, one obtainsa fromsuch a fitting s0 = 1.6ω and E0 = 0.9ω. Adding the third order terms improvesthe result: s0 = 1.75ω and E0 = 0.95ω. In our case one can substitute into thesum rule the all-order result. The remaining ε-dependence is then entirely dueto the use of an approximate model for the higher states. For sufficiently smallvalues of ε all the curves give the exact value E0 = 2ω. The flattest curves arethose with s0 ∼ 2ω.

The second step is to extract the ψ(0) value from the original sum rule(30)

|ψ0(0)|2 = eE0/ε

ε

2(1− e−s0/ε)− ω2

12ε+

7720

ω4

ε3− 31

30240ω6

ε5+ . . .

. (33)

The analysis goes in a similar way. If one takes the third-order value E0 =0.95ω, one obtains |ψ0(0)|2 ∼ 0.9ω. In the ε→∞ limit one obtains the localduality relation

|ψ0(0)|2LD =s0

2. (34)

So, the “exact value” for the duality interval s0 is 2ω.Summarizing, using the sum rule method one can determine the param-

eters of the lowest resonance with ≈ 10% accuracy. The main source of theerror is the use of a rather rough model for the spectral density of the higherstates. It should be noted that in QCD situation is better: normally only thefirst resonance is narrow, while higher states are broader and broader, so thatapproximating them by free quark functions does not lead to large errors.

aPlotting E0 for various values of s0 is absolutely straightforward, and we encourage thereaders perform the fitting as an excercise.

9

2.2 Sum rules for form factors of the oscillator states.

To calculate the form factors

Fkl(Q) =∫ψ∗k(x)ψl(x)eiQx d2x (35)

within the sum rule approach, one should construct a similar integral usingthe Green functions (13)

G(x, 1/iε | 0, 0 ) =∞∑k=0

ψ∗k(x2)ψk(x1)e−Ek/ε . (36)

This gives the amplitude

M(ε1, ε2, Q2) =∫G(x, 1/iε1 | 0, 0 )G∗(x, 1/iε2 | 0, 0 )eiQx d2x

=∞∑

k,l=0

ψ∗l (0)ψk(0)[∫

eiQxψ∗k(x)ψl(x)d2x

]e−Ek/ε1−El/ε2

=∞∑

k,l=0

ψ∗l (0)ψk(0)Fkl(Q2)e−Ek/ε1−El/ε2 (37)

One can write this amplitude in the (double) spectral representation:

M(ε1, ε2, Q) =1π2

∫ ∞0

∫ ∞0

e−s1/ε1−s2/ε2ρ(s1, s2, Q2)ds1ds2 (38)

where the spectral function is the sum of double delta-functions:

ρ(s1, s2, Q2) = π2

∞∑k,l=0

ψ∗l (0)ψk(0)Fkl(Q) δ(s1 −Ek) δ(s2 −El). (39)

In the free case one has

Gfree(x, 1/iε | 0, 0) =mε

2πe−mx

2ε/2 =∫

d2k

(2π)2eikxe−k

2/2mε (40)

for the Green function and

Mfree(ε1, ε2, Q2) =∫

d2k

(2π)2e−k

2/2mε1−(k+Q)2/2mε2

=(m

2π

) ε1ε2ε1 + ε2

exp(− Q2

2m(ε1 + ε2)

)(41)

10

for the amplitude. The free spectral function ρfree(s1, s2, Q2) can be easily

extracted:

ρfree(s1, s2, Q2) = π2

∫d2k

(2π)2δ

(s1 −

k2

2m

)δ

(s2 −

(k +Q)2

2m

)=θ(|√

2ms1 +√

2ms2| ≥ |Q| ≥ |√

2ms1 −√

2ms2|)√4s1s2 − (s1 + s2 −Q2/2m)2

.(42)

The restriction on s1, s2 is in fact the triangle relation

|k1|+ |k2| ≥ |Q| ≥ | |k1| − |k2| |

between the initial momentum k1 (|k1| =√

2ms1), the momentum transfer Qand the final momentum k2 = (k1 +Q) (|k2| =

√2ms2).

The oscillator Green function is also known:

Gosc(x, 1/iε | 0, 0) =mω

2π sinh(ω/ε)exp

(−mω~x

2 cosh(ω/ε)2 sinh(ω/ε)

)(43)

and using it, one can calculate the amplitude

M(ε1, ε2, Q2) =mω

2π sinh(ω/ε1 + ω/ε2)exp

(− Q2

2mωsinh(ω/ε1) sinh(ω/ε2)

sinh(ω/ε1 + ω/ε1)

).

(44)Now, representing M as

M(ε1, ε2, Q2) =mω

π

∞∑k=0

e−(2k+1)ω(1/ε1+1/ε2) exp(− Q2

4mω

)exp

(Q2

4mω

[1− (1− e−ω/ε1)(1− e−ω/ε2)

(1− e−ω/ε1e−ω/ε2)

])(45)

and expanding M in powers of e−ω/ε1 ≡ E1 and e−ω/ε2 ≡ E2 one can extractthe form factors Flk(Q2):

M(ε1, ε2, Q2) =mω

π

∞∑k=0

E2k+11 E2k+1

2 exp(− Q2

4mω

)

1 +Q2

4mω

[1− (1− E1)(1− E2)

1− E1E2

]+

12

(Q2

4mω

)2

[. . .]2 + . . .

(46)

identifying them as the factors in front of the Ek1 E l2 terms:

F00(Q2) = exp(− Q2

4mω

),

11

F01(Q2) = F10(Q2) =Q2

4mωexp

(− Q2

4mω

),

F11(Q2) =(

1− Q2

4mω

)2

exp(− Q2

4mω

)(47)

It is instructive to write down the amplitude M(ε1, ε2, Q2) for Q2 = 0:

M(ε1, ε2, Q2 = 0) =mω

2π sinh(ω/ε1 + ω/ε2)=mω

π

∞∑k=0

E2k+11 E2k+1

2 . (48)

It is easy to notice that• There are no terms with k 6= l: only the diagonal transitions are allowed inthis limit• The numerical factor is just the same as in Eq. (14) for Mosc(ε). This meansthat Fkk(0) = 1 for all k, just as required by charge conservation.

Next step is to get the perturbative expansion in powers of ω :

M(ε1, ε2, Q2) =mε1ε2

2π(ε1 + ε2)

[1− ω2

6

(1ε1

+1ε2

)2

+ . . .

]

× exp

−Q2

2m

1ε1

(1 + 16ω2

ε21+ . . .) 1

ε2(1 + 1

6ω2

ε22+ . . .)

( 1ε1

+ 1ε2

)(1 + ω2

6 ( 1ε1

+ 1ε2

)2 + . . .)

=

mε1ε22π(ε1 + ε2)

[1− ω2

6

(1ε1

+1ε2

)2

+7ω4

360

(1ε1

+1ε2

)4

+ . . .

]

exp(− Q2

2m(ε1 + ε2)

)[1 +

Q2ω2

6mε1ε2(ε1 + ε2)

− Q2ω4[(ε1 + ε2)2 + 2ε1ε2]90mε31ε32(ε1 + ε2)

+Q4ω4

72m2ε21ε22(ε1 + ε2)2

+ . . .

]= Mfree(ε1, ε2, Q2)

1 +

∞∑k=1

(ω2)kϕk(ε1, ε2, Q2)

(49)

The sum rule now reads

|ψ0(0)|2F00(Q2)e−E0/ε1−E0/ε2 + “higher states”

= Mfree(ε1, ε2, Q2)

1 +O

(ω2

ε2

)+O

(ω4

ε4

)+ . . .

(50)

12

To proceed further, one should take some ansatz for the higher statescontribution. The ansatz similar to that used earlier is to assume that

“higher states” = “free states”

outside the square (0, s0)⊗ (0, s0). Then the sum rule has the form

|ψ0(0)|2F00(Q2)e−E0/ε1−E0/ε2

=1π2

∫ s0

0

ds1

∫ s0

0

ds2 ρfree(s1, s2, Q

2)e−s1/ε1−s2/ε2

+O(ω2

ε

)+O

(ω4

ε3

)+ . . . . (51)

In the ε1, ε2 →∞ limit one obtains the local duality relation:

|ψ0(0)|2F00(Q2) =1π2

∫ s0

0

∫ s0

0

ρfree(s1, s2, Q2)ds1ds2 (52)

with ρfree(s1, s2, Q2) given by Eq. (42):

ρfree(s1, s2, Q2) =

14

∫d2k δ

(s1 −

k2

2m

)δ

(s2 −

(k +Q)2

2m

). (53)

Using the two expressions given above, one can write

|ψ0(0)|2FLD00 (Q2) =1

(2π)2

∫d2k θ

(k2

2m< s0

)θ

((k +Q)2

2m< s0

). (54)

This integral can be easily calculated if one notices that it defines the overlaparea of two circles having the same radius R2 = 2ms0 , with their centersseparated by the distance equal to Q:

|ψ0(0)|2FLD00 (Q2) =ms0

π2

[cos−1(κ)− κ

√1− κ2

]θ(Q2 < 8ms0) , (55)

where κ = Q/√

8ms0. Recall now that from the local duality relation for M(ε)we had

|ψ0(0)|2LD =ms0

2π(56)

and this gives the result

FLD00 (Q2) =2π

[cos−1(κ)− κ

√1− κ2

]θ(Q2 < 8ms0) (57)

13

having the correct normalization FLD00 (Q2) = 1.The local duality prediction calculated for the “exact value” s0 = 2ω of

the duality interval

FLD00 (Q2)|s0=2ω =2π

[cos−1

(Q√

16mω

)− Q√

16mω

√1− Q2

16mω

]θ(Q2 < 16mω) (58)

is very close to the exact form factor

F exact00 (Q2) = exp(− Q2

4mω

). (59)

There is a very simple physics behind the local duality. One should justcompare the local duality prescription for the form factor

FLD00 (Q2) =1

|ψ0(0)|2∫

d2k

(2π)2θ

(k2

2m< s0

)θ

((k +Q)2

2m< s0

)(60)

and the exact formula

F00(Q2) =∫

d2k

(2π)2ψ0(k)ψ∗0(k +Q) (61)

to realize that the local duality, in fact, is equivalent to using a model form forthe wave function in the momentum space:

ψLD0 (k) =1

|ψ0(0)| θ(k2 < 2ms0) (62)

or, taking s0 = 2ω:

ψLD0 (k) =√

π

mωθ(k2 < 4mω), (63)

to be compared with the exact wave function

ψ0(k) = 2√

π

mωexp

(− k2

2mω

). (64)

It is easy to establish that the model wave function correctly reproduces thesimplest integral properties of the exact wave function:∫

ψexact(k)d2k =∫ψLD(k)d2k (65)∫

|ψexact(k)|2d2k =∫|ψLD(k)|2d2k. (66)

14

3 Condensates ant Operator Product Expansion

3.1 Correlators

The basic object in the quantum mechanical example was the Green functionG(x1, t1 |x2, t2 ) describing the propagation of a particle from one point toanother. In fact, it was sufficient to analyze the probability amplitude forthe partricle to remain at the same spatial point after some imaginary timeelapsed:

M(ε) = G( 0, 0 | 0, 1/iε ) . (67)

The question is what should be its analog in the quantum field theory?Recall, that the motion in the external potential corresponds to a two-particlesystem with one of the particles being infinitely massive. If this particle moves,the other moves together with it. So, the Green function described really acollective motion of two particles from one point to another. In quantum fieldtheory this corresponds to the Green-like functions

〈0|T (j(x) j(y) )| 0〉

normally referred to as correlators, with the currents j(x) being the productsof fields at the same point, e.g. Jµ(x) = q(x)γµq(x) . Due to the translationinvariance

〈0|T (j(x) j(y) )| 0〉 → 〈0|T (j(x− y) j(0) )| 0〉it is always possible and convenient to take one point at the origin. In QFT, it isalso more convenient to consider correlators in the momentum representation:

Π(q) =∫eiqx〈0|T (j(x) j(0) )| 0〉 d4x.

In the lowest order the correlator is given by the product of the relevant prop-agators, but there are also radiative corrections, which one can calculate inperturbative QCD, at least in the region where q is spacelike q2 < 0 and largeenough.

Next step is to relate Π(q) to a sum over the physical states. This isperformed via the dispersion relation

Π(q2) =1π

∫ ∞0

ρ(s)s− q2

ds+ “subtractions”.

The subtraction terms appear if the spectral density is such that the dispersionintegral diverges. If, say, ρ(s) ∼ sn, one needs (n+ 1) subtractions:

Π(q2) =1π

∫ ∞0

ρ(s)s− q2

ds

15

− 1π

∫ ∞0

ds ρ(s) 1s− λ2

+q2 − λ2

(s− λ2)2+ . . .+

(q2 − λ2)n

(s− λ2)n+1

+ Π(λ2) + (q2 − λ2)Π′(λ2) + . . .+(q2 − λ2)n

n!Π(n)(λ2). (68)

After collecting all the terms containing ρ(s), one observes that ρ(s) is dividedby (s−q2)(s−λ2)n+1, and the integral converges. Note, that all the subtractionterms are polynomials of a finite order in Q2 (from now on we use the notationq2 = −Q2, with Q2 > 0).

3.2 Borel transformation

Note, that in the oscillator studies, we dealt with the exponentially weightedsum:

M(ε) =1π

∫ ∞0

e−s/ερosc(s) ds. (69)

We observed then that it has a nice property that for ε values of an order ofω, both the contributions of the higher states were strongly damped and theperturbative expansion for M(ε) was converging fast enough. With the powerweight implied by the dispersion relation, the suppression of the higher statescontribution is not so effective.

However, it is rather easy to transform the original dispersion integral intothe exponentially weighted one. To this end, one should apply the so-calledBorel transformation, formally defined by the formula1

Φ(M2) = limQ2,n→∞Q2/n=M2

1(n− 1)!

(Q2)n[− d

dQ2

]nΠ(Q2). (70)

It is easy to check that applying the above procedure to 1/(s+Q2) one reallyobtains an exponential. Indeed,[

− d

dQ2

]n 1s+Q2

=n!

(s+Q2)n+1. (71)

Multiplying the result by (Q2)n

(n−1)! one obtains

n(Q2)n

(s+Q2)n+1=

1Q2/n

1(1 + s/Q2)n+1

=1M2

1(1 + s/nM2)n+1

. (72)

Finally, taking the n→∞ limit one obtains 1M2 e

−s/M2. This means that

Φ(M2) ≡ B(Q2 →M2) Π(Q2) =1π

∫ ∞0

ds

M2ρ(s)e−s/M

2+ “nothing” (73)

16

where “nothing” means that all the subtraction terms have disappeared, sinceno polynomial can survive after an infinite number of differentiations. Thus,the function Φ(M2) is the QFT analog of M(ε).

In many cases one can perform the Borel transformation using the formula

B(Q2 →M2)e−AQ2 = δ(1−AM2). (74)

As an illustration, let us reproduce the result obtained above :

B1

s+Q2= B

∫ ∞0

e−α(s+Q2) dα

=∫ ∞

0

e−αsδ(1− αM2) dα =1M2

e−s/M2. (75)

Other functions, the Borel transforms of which we will need, are those havingthe power behavior in 1/Q2 . In this case we have

B1

(Q2)n= B

∫ ∞0

e−αQ2 αn−1

(n− 1)!dα

=∫ ∞

0

δ(1− αM2)αn−1

(n− 1)!dα =

1(n− 1)!(M2)n

. (76)

Thus, a power expansion for Φ(M2) over 1/M2 converges much faster thanthe original expansion of Π(Q2) over 1/Q2, since each term is now suppressedfactorially. This looks precisely like the Borel improvement of a series, andthat is why Shifman, Vainshtein and Zakharov1 referred to the transformationthey introduced as “the Borel transformation”. The representation for Φ(M2)has the form of the Laplace transformation, and that is why some people callthe sum rules obtained in this way “the Laplace transform sum rules”. Infact, there is no contradiction: Φ(M2) is the Borel transform of Π(Q2) and theLaplace transform of ρ(s).

To summarize, the Borel transformation produces two improvements:• higher state contributions in the spectral representation for Φ(M2) are ex-ponentially suppressed,• its power series expansion in 1/M2 is factorially improved.

3.3 Heavy-light system and the m→∞ limit

Using the basic formula, it is rather easy to establish11 that in the limit whenone of the particles becomes infinitely massive, the QFT function Φ(M2) con-verts into its quantum-mechanical analog. Consider the lowest-order diagram.

17

The propagator of the heavy particle is

1(q + k)2 −m2

=1

q2 −m2 + 2qk + k2.

In the m → ∞ limit all the bound state masses are equal to this mass plussome finite quantity. So, it makes sense to subtract this infinite constant. Tothis end one can represent the total momentum q as

q = mv + l

where v is the four-velocity. Then

q2 −m2 = 2m(vl) + l2

The scalar product (vl) ≡ E just has the meaning of the energy variable. Thereis only one more O(m) term in the propagator:

2(qk) = 2m(vk)1 +O(1/m) .

With this accuracy, one can write down the relevant Feynman integral as

I(m, v,E) =∫

d4k

2m(E + (vk))S(k) ,

where S(k) is the propagator of the light particle. Now, applying the Boreltransformation B(E → ε) which converts 1/(E + (vk)) into 1

ε e−(vk)/ε we get

M(m, v, ε) ≡ B(E → ε)I(m, v,E) =1

2mε

∫e−(vk)/εS(k)d4k.

It is easy to realize that the integral above is just the light particle propagatorin the configuration space

M(m, v, ε) =1

2mε〈ϕ(0)ϕ(iv/ε)〉

i.e., just the Green function describing propagation in the imaginary time,since the velocity vector vµ, in the rest frame of the heavy particle, is orientedexactly in the time direction.

Thus, we have demonstrated that the Borel transformed correlators arejust the QFT generalization of the imaginary time Green functions we studiedin the quantum-mechanical examples.

18

3.4 Perturbative vs nonperturbative contributions in QCD

From the technical point of view, the strategy now is• to calculate the correlator Π(Q2) in the deep spacelike region where one canrely on the asymptotic freedom of QCD,• apply the Borel transformation to the resulting expression to get a 1/M2

expansion for Φ(M2),• finally, using the dispersion representation for Φ(M2), to construct a sumrule.

The question is, what are the contributions analogous to the ω2/ε2 correc-tions?

The standard idea about the QCD potential V (r) is that it consists of aCoulomb-like part ∼ αs(1/r2Λ2) calculable perturbatively at short distancesand a nonperturbative ∼ r long distance “confining” part:

V (r) = V “Coulomb”(r) + V conf (r)

The magnitude of the “Coulomb” effects is determined by the QCD run-ning coupling constant

αs(Q2) =4π

9 log(Q2/Λ2)+ . . . .

The standard modern estimate for the value of the QCD scale Λ is

Λ ∼ 200MeV .

As we will see, the parameters that really appear in QCD sum rules havedimension of (mass)2 , and one should compare

Λ2 ∼ 0.04 GeV2

to the typical hadronic scales like ρ and nucleon masses (squared):

m2ρ ≈ 0.6 GeV2, m2

N ≈ 0.9 GeV2.

To get such scales from Λ2 one needs unnaturally large numerical factors ofan order of 25 or even 100. So, it is most unlikely that the O(αs)-correctionscan generate scales of an order of 1 GeV and , hence, the “Coulomb” part isirrelevant to the formation of the hadronic spectrum. Our problem now is totake into account the effects due to the long-range nonperturbative part of theQCD potential.

Our study of the quantum-mechanical oscillator (which is the simplestexample of a system with confinement) demonstrated that the propagation

19

amplitude in the presence of the potential is different from the free Green func-tion. Furthermore, we observed that the difference vanishes at short distancesand that one can calculate exact Green function perturbatively, expanding inpowers of the oscillator potential. In QCD, with two components of V (r) onecan imagine a double series, both in V “Coulomb”(r) and V conf (r). There is noproblem to arrange an expansion in V “Coulomb”(r) , − this will be just the stan-dard perturbation theory in the QCD coupling constant. On the other hand,the confining potential V conf (r) in QCD is not even known. Moreover, thewidespread belief is that it is determined by some essentially nonperturbativeeffects, i.e., by those which cannot be seen in perturbation theory.

In such a situation, a possible way out is to proceed as follows:• to construct perturbation expansion in terms of quark and gluon propagators(this amounts to taking into account only the Coulombic part of the potential),• to postulate that quark and gluon propagators are modified by the long-rangeconfinement part of the QCD potential; but the modification is soft in a sensethat at short distances the difference between exact and perturbative (free-field) propagators vanishes.

To formalize this statement, one can write the exact propagator Dexact(x)as a vacuum average of a T-product of fields in the exact vacuum Ω (corre-sponding to a confining potential)

Dexact(x) = 〈Ω|T (ϕ(x)ϕ(0))|Ω〉.

According to the Wick theorem, one can write the T−product as the sum

T (ϕ(x)ϕ(0)) = ϕ(x)ϕ(0)︸ ︷︷ ︸+ : ϕ(x)ϕ(0) :

of the “pairing” and the “normal” product. The “pairing” is just the expecta-tion value of the T - product over the perturbative vacuum

ϕ(x)ϕ(0)︸ ︷︷ ︸ = 〈0|T (ϕ(x)ϕ(0))|0〉.

i.e., the perturbative propagator. By this definition, the normal product: ϕ(x)ϕ(0) : vanishes if averaged over the perturbative vacuum:

〈0| : ϕ(x)ϕ(0) : |0〉 = 0.

Thus, our assumption that Dexact(x) 6= Dpert(x) is equivalent to the statement

〈Ω| : ϕ(x)ϕ(0) : |Ω〉 6= 0,

which is the starting point to calculating power corrections in QCD. To simplifythe notation, in what follows we will write the lhs of the above equation as〈ϕ(x)ϕ(0)〉.

20

3.5 Condensates

In the oscillator case the analog of 〈ϕ(x)ϕ(0)〉 is the difference between theexact Green function

Gosc( 0, 0 | 0, τ/i) =mω

2π sinh(ωτ)

and the free Green function

Gfree( 0, 0 | 0, τ/i) =m

2πτ

where τ ≡ 1/ε is just the imaginary time variable. Note, that both oscillatorand free Green function are singular for τ → 0, but the difference

Gosc( 0, 0 | 0, τ/i)−Gfree( 0, 0 | 0, τ/i)

=m

2π

[−1

6ω2τ +

7360

ω4τ3 − 3115120

ω6τ5 + . . .

](77)

is regular at that point, and one can even expand the difference into the Taylorseries in (ωτ)2. The expansion is possible essentially because the confiningpotential is regular (and even vanishing) at τ = 0.

In the QCD sum rule approach it is also assumed that the confinementeffects are sufficiently soft to allow for the Taylor expansion of 〈ϕ(x)ϕ(0)〉 atx = 0:

〈ϕ(0)ϕ(x)〉 =∞∑n=0

xµ1 . . . xµn

n!〈ϕ(0)∂µ1 . . . ∂µnϕ(0)〉

= 〈ϕϕ〉+ xµ〈ϕ∂µϕ〉 +xµ1xµ2

2〈ϕ∂µ1∂µ2ϕ〉+ . . . (78)

This is, in fact, the expansion of the nonlocal object 〈ϕ(0)ϕ(x)〉 over thevacuum matrix elements of the local composite operators. Not all operatorsreally contribute to the expansion above. Using the Lorentz covariance, it istrivial to find that

〈ϕ∂µϕ〉 = 0,

〈ϕ∂µ1∂µ2ϕ〉 =14gµ1µ2〈ϕ∂2ϕ〉, (79)

etc., so that finally one arrives at the expansion in x2:

〈ϕ(0)ϕ(x)〉 =∞∑n=0

(x2

4

)n 1n!(n+ 1)!

〈ϕ(∂2)nϕ〉 . (80)

21

Thus, the modification of the propagator by the nonperturbative effects is nowparametrized by the matrix elements of the composite operators like 〈ϕ(∂2)nϕ〉.The examples in QCD are• 〈qq〉 referred to as the quark condensate,• 〈qD2q〉, characterizing the average virtuality of the vacuum quarks,• the gluon condensate 〈GaµνGaµν〉 , etc. Here Dµ ≡ ∂µ− igAµ is the covariantderivative and Gµν = (i/g)[Dµ, Dν ] is the gluonic field strength. Note, thatonly gauge invariant composite operators should appear in QCD, i.e. each ∂µmust be accompanied by the relevant Aµ. How this happens, will be discussedlater.

3.6 Operator product expansion

It is instructive to analyze the above expansion in the momentum representa-tion

〈ϕ(0)ϕ(x)〉 ≡∫eikxD(k) d4k =

∞∑n=0

(x2

4

)n 1n!(n+ 1)!

∫(−k2)nD(k) d4k.

(81)Note, that existence of 〈ϕ(0)ϕ(0)〉 implies that D(k), the nonperturbative partof the propagator, vanishes for large momenta faster than k−5, while exis-tence of 〈ϕ(0)∂2ϕ(0)〉 implies that D(k) < k−7, etc. All matrix elements〈ϕ(0)(∂2)nϕ(0)〉 exist only if, for large momenta, the function D(k) vanishesfaster than any power of 1/k2, say, like an exponential. Thus, the functionD(k) is concentrated in the region of small momenta. This corresponds tothe basic assumption that at large momenta one should enjoy the asymptoticfreedom.

In other words, the exact propagator is represented as a sum of the pertur-bativeO(1/k2) part and the nonperturbative (“condensate”) part that vanishesfast when the momentum k increases. Using such a decomposition, we obtaina modified diagram technique for the correlator functions. Consider, e.g., thelowest order diagram corresponding to the product of two propagators. Afterthe decomposition it produces four diagrams (see Fig.3), where the dashedlines correspond to the nonperturbative part of the propagators.

The first term (Fig.3a) is just the ordinary purely perturbative diagramhaving ∼ log(q2) behavior for large q2:∫

d4k

k2(q − k)2∼ log(q2).

Two other diagrams (Figs.3b,c), as we will see below, behave like 1/q2. Thisbehavior is completely determined by that of the perturbative propagator: the

22

k k k k

q−k q−k q−kq−kq q qq

a) b) c) d)

Figure 3: Operator product expansion for scalar one-loop diagram.

whole momentum flows through the perturbative line where its propagation isnot suppressed. In the last diagram (Fig.3d), all possible momentum flows arestrongly suppressed, e.g., if D(k), the nonperturbative part of the propagator,vanishes exponentially at large k, so does this contribution also.

Now, let us analyze the contribution of the second diagram:

I〈ϕϕ〉(q2) =∫D(k2)

d4k

(q − k)2. (82)

As we discussed earlier, the function D(k2) is concentrated in the region ofsmall momenta. So it makes sense to expand the perturbative propagator1/(q − k)2 in k:

1(q − k)2

=1q2

∞∑n=0

2nqk(n)

(q2)nθ(k2 < q2) +

1k2

∞∑n=0

2nqk(n)

(k2)nθ(q2 < k2) (83)

whereqk(n) ≡ qµ1 . . . qµnkµ1 . . . kµn

is the product of two symmetric

O...µi...µj ... = O...µj ...µi...

and tracelessO...µi...µj ...gµiµj = 0

tensors formed from the vectors q and k, respectively. The symmetric-tracelesstensors have a very useful property:∫

F (k2)kµ1 . . . kµn d4k = 0

23

for n 6= 0, because it is impossible to construct a nontrivial symmetric-tracelesscombination using only the metric tensors gµiµj . To do this, one should haveat least one vector, e.g.,

pµ1pµ2 = pµ1pµ2 −14p2 gµ1µ2 .

Thus, in our case

I〈ϕϕ〉(p2) ≡∫D(k2)

d4k

(q − k)2

=1q2D(k2)θ(k2 < q2)

∫d4k +

∫D(k2)

d4k

k2θ(q2 < k2)

=1q2〈ϕϕ〉 + “quickly vanishing contribution”, (84)

where the “quickly vanishing” term is given by the integral over large momenta:

“quickly vanishing contribution” =∫ (

1k2− 1q2

)D(k2) θ(k2 > q2) d4k

where the function D(k2) is strongly suppressed.Thus, it is very easy to calculate this contribution: the whole external

momentum flows through the perturbative line, so one should just multiplythe perturbative propagator 1/q2 by the condensate 〈ϕϕ〉 factor. One canwonder, however, why the higher condensates (containing the ∂2) did not ap-pear? To answer this question, it is useful to analyze this contribution in theconfiguration representation. The original diagram decomposition now lookslike

D(x, 0)D(0, x) = Dpert(x, 0)Dpert(0, x) +Dpert(x, 0)〈ϕ(0)ϕ(x)〉+ Dpert(0, x)〈ϕ(x)ϕ(0)〉 + 〈ϕ(0)ϕ(x)〉〈ϕ(x)ϕ(0)〉, (85)

and its second term can be written (modulo trivial factors like 2π) as

〈ϕ(0)ϕ(x)〉/x2.

Now, using the x2-expansion for the 〈ϕ(0)ϕ(x)〉- factor, we obtain

1x2〈ϕ(0)ϕ(x)〉 =

1x2〈ϕϕ〉 +

1x2

x2

8〈ϕ(0)ϕ(x)〉 +O(x2) + . . . . (86)

In the momentum representation, the first term just corresponds to the 〈ϕϕ〉/q2

contribution, while the second is proportional to δ4(q) and can be discarded

24

since the external momentum q is assumed to be large. Higher terms are pro-portional to the derivatives of δ4(q) , and also can be ignored for large q. Itis easy to realize, that the sum of all these terms just produces the “quicklyvanishing contribution” discussed above. Only the terms singular for x2 = 0(in the configuration representation) can produce a power-like behavior in themomentum representation.

The expansion we discussed above is a simplified illustration of the operatorproduct expansion12

T (j(x)j(0)) =∑n=0

Cn(x2)On(0). (87)

It represents the T -product of two currents j(x), j(0) as a sum of the local(composite) operators of increasing dimension (in units of mass). The coeffi-cient functions Cn(x2) characterizing the contribution of a particular operatorare ordered in their singularity at x2 = 0. From a simple dimensional analysisit follows that there should be a simple relation between singularity of the co-efficient function and the dimension of the relevant composite operator; e.g.,if

Cn(x2) ∼ 1(x2)N0−n

thenOn ∼ (µ2)d0+n,

so that after the Fourier transform one arrives at an expansion in powers ofµ2/q2. There are, of course, more complicated contributions into the OPE,corresponding to more complicated original diagrams.

Similar decomposition of propagators into perturbative and nonperturba-tive parts can be also performed in the QCD case. At this level of discussion,the only difference is that in QCD one has to deal with two types of fields:quarks q, q and gluons A. Consider the decomposition of a two-loop diagram(see Fig.4).• The first diagram (Fig.4a) corresponds to the purely perturbative contribu-tion (unity operator in the OPE language),• the second diagram (Fig.4b) (in fact there are two such diagrams) corre-sponds to the 〈qAq〉-type contribution,• the third (Fig.4c) is evidently proportional to the 〈AA〉 vacuum average,accompanied by the one-loop coefficient function,• the fourth (Fig.4d) diagram (there is also another such diagram) brings inthe four-quark vacuum average 〈qqqq〉.

25

qqqq

a) b) c) d)

Figure 4: Operator product expansion for QCD two-loop diagram.

4 Operator Product Expansion: Gauge Invariance and RelatedProblems

Looking at the operators discussed in the previous section, one can noticethat some of them, e.g., 〈q∂µq〉, 〈qAµq〉, 〈AµAν〉, etc., do not satisfy a veryimportant requirement: they are not gauge invariant. And the product ofgauge invariant operators in a gauge theory should have an expansion in gaugeinvariant operators like 〈qDµq〉 and 〈GaµνGaµν〉.

4.1 Notations

Let us introduce now the necessary notations. To begin with,

Dµ = ∂µ − igAµ

is the covariant derivative in the quark (fundamental) representation of theSU(3)c gauge group:

Aµ = Aaµτa,

with matrices (τa)AB simply related to the Gell-Mann (3× 3) λ-matrices

τa ≡ λa/2.

The indices A,B correspond to 3 possible quark colors:A,B = 1, 2, 3. Thegluons have 8 colors described by the a-indices: a = 1, . . . 8. The covariantderivative

Dµ = ∂µ − igAµacting on the gluonic field contains the (8× 8) σ-matrices

Aµ = Aaµσa

26

characteristic for the gluonic (adjoint) representation of the SU(3) color group:

(σa)bc = −ifabc,

where fabc are the structure constants of the group:

[τa, τb] = ifabcτc.

The gluonic field strength tensor

Gaµν = ∂µAν − ∂νAµ + gfabcAbµAcν

can be related to the commutator of the covariant derivatives:

Gµν =1ig

[Dµ, Dν ] (88)

Gµν =1ig

[Dµ, Dν ]. (89)

4.2 External field method

A rather innocent-sounding statement that only the gauge invariant operatorsshould appear in the operator product expansion implies that there are nu-merous nontrivial correlations between the coefficient functions of non-gauge-invariant operators originating from completely different diagrams. For ex-ample, the operator 〈q∂µq〉, appears when one expands the nonlocal operator〈q(0)q(x)〉 in the simplest combination (see Fig.5a)

〈q(0)S(−x)q(x)〉,

while the operator 〈qAµq〉 necessary to form a gauge-invariant sum, resultsfrom a more complicated combination(see Fig.5b)∫

〈q(0)S(−y)γµ(ig)Aµ(y)S(y − x)q(x)〉 d4y.

In both cases, the perturbative propagators (in configuration space) are writtenexplicitely, while the nonperturbative parts, in accordance with our previousdiscussions, are denoted by vacuum averages of the relevant normal productsof fields. By an explicit integration one can establish that∫

S(−y)γµS(y − x) d4y = −xµS(−x) ,

27

+y

0x x x 00x 0++

d)c)b)a)

Figure 5: Summation over gluons inserted into hard quark propagator.

and, hence, the coefficient functions of the two operators are just those neces-sary to form the gauge invariant combination.

However, this is definetely not the method one would like to use in morecomplicated situations. In fact, it is much easier to analyze the sum of all thecombinations differing only by the number of the gluons going into vacuum(see Fig.5).

One can write it as13,14 〈q(0)S(0, x;A)q(x)〉 , where S(0, x;A) has themeaning of the perturbative quark propagator in the (external) vacuum gluonicfield A. It satisfies the Dirac equation[

iγµ(∂

∂xµ− igAµ(x)) −m

]S(x, y;A) = −δ4(x− y). (90)

Solving it by using perturbation theory in A, one gets back the original dia-grammatic expansion.

The function S(x, y;A) is not gauge invariant. From the good old electro-dynamics it is known that propagation of a charged particle in the externalfield results in a phase factor given by the exponential of a line integral. In theQCD case it looks like

P (x, y;A; C) = P exp(ig∫CAµ(z) dzµ),

where P means that the color matrices should be ordered along the path Cconnecting the points x and y. Furthermore, it is known that the combination

〈q(y)P (y, x;A; C)q(x)〉

is gauge invariant (though path-dependent). This gives the idea how one canconstruct the operator product expansion in an explicitly gauge invariant way.Let us try the ansatz

S(x, y;A) = E(x, y;A)S(x, y;A) (91)

28

where E(x, y;A) is the P -exponential for the straight-line path connecting thepoints x and y:

E(x, y;A) = P exp[ig(xµ − yµ)

∫ 1

0

Aµ(z) dt]z=y+t(x−y)

.

It is straightforward to derive that the original Dirac equation is satisfiedonly if the function S(x, y;A) is a solution to the modified Dirac equation[

iγµ(∂

∂xµ− igAµ(x; y))−m

]S(x, y;A) = −δ4(x− y) (92)

that differs from the original one only by the change

Aaµ(x)→ Aaµ(x; y) = (xν − yν)∫ 1

0

Gbνµ(z) Eba(z, y) tdt. (93)

Here, z = y + t(x − y) and E is a straight-line-ordered exponential with theA-fields multiplied by the σ-matrices of the gluonic representation. It appearsafter one commutes the E-exponentials through a τ -matrix:

(τa)ABEBC(z, y) = EAB(z, y)(τb)BCEab(z, y). (94)

The commutation rule is based on the well-known formula

eABe−A = B + [A,B] +12!

[A, [A,B]] + . . . (95)

and the relation[τa, τb] = −(σa)bcτc. (96)

Surprizingly enough, the presence of the E-factor only simplifies the situ-ation, because the derivatives in the Taylor expansion of the GE-combinationare just the covariant ones:

Gbνµ(z) Eba(z, y) =∞∑n=0

1n!Gaνµ;µ1...µn(y)(z − y)µ1 . . . (z − y)µn . (97)

As a result, the new field Aaµ(x; y) can be expressed in terms of the gluon fieldstress tensor Gµν and its covariant derivatives (taken, of course, in the gluonicrepresentation):

Aaµ(x; y) =∞∑n=0

1(n+ 2)n!

(x− y)ν(x− y)µ1 . . . (x− y)µn Gaνµ;µ1...µn(y). (98)

29

This means that the function S(x, y;A) depends on the gluonic field onlythrough the gluon field stress tensor Gµν and its covariant derivatives. Solvingthe equation for S(x, y;A) perturbatively, one obtains the gA-expansion thathas the same structure as the original gA-expansion. A very important dif-ference is that the resulting operators have an explicitly gauge invariant form.For example, taking the lowest (zero) order term one obtains the contribution

〈q(y)S(y − x)E(y, x;A)q(x)〉 (99)

containing the gauge-invariant operator

〈q(y)E(y, x;A)q(x)〉 =∞∑n=0

1n!

(x− y)µ1 . . . (x− y)µn〈qDµ1 . . . Dµnq〉|y. (100)

We used the direct straight-line path from x to y in our ansatz. Usually, itis more convenient to use the path formed by two straight lines: from x to somefixed point z0 (usually taken at the origin, z0 = 0) and then from z0 to y. Theonly change then will be that quark and gluon fields q(ξ), A(ξ) will be finallyaccompanied by the E(ξ, z0;A) exponentials, so that a covariant expansionswill be obtained if one expands at the fixed point z0. Using translation in-variance, one can always substitute matrix elements 〈q(z0) . . . G(z0) . . . q(z0)〉,resulting from such a procedure, by 〈q(0) . . .G(0) . . . q(0)〉. Less trivial is thedisappearance of the z0-dependence from the coefficient functions. One canenjoy it only if the calculations had no errors, and after summation over alldiagrams (of the modified expansion) producing the same operator. The latterobservation suggests that z0 works as a gauge parameter. Indeed, it is easy tonotice that the relevant P -exponentials

E(x, z0; a) = P exp[ig(xµ − zµ0 )∫ 1

0

Aµ(z) dt]|z=z0+t(x−z0) (101)

disappear (are equal to 1) if one imposes the gauge condition15,16,17,18,14

(xµ − zµ0 )Aµ(x) = 0. (102)

Furthermore, the “calligraphic” field A in this (Fock-Schwinger or fixed point)gauge coincides with the ordinary one A = A(G), i.e., in this gauge not onlyGµν can be expressed in terms of Aµ, but also the vector-potential Aµ can berepresented as a (nonlocal) functional of the field strength tensor Gµν . Thismeans that the Fock-Schwinger gauge is a physical gauge. Another example isthe axial gauge19

nµAµ(x) = 0 (103)

30

with nµ being a fixed vector. In this gauge one can also express19 Aµ in termsof Gµν :

Aµ(x) = nν∫ ∞

0

Gνµ(x+ tn) dt. (104)

Imposing the Fock-Schwinger gauge condition, one can forget about theexponentials. One should remember, however, that the “ordinary” derivativesin this gauge should be treated as the covariant ones, e.g., they do not com-mute, etc. To illustrate the typical tricks and conventions, let us consider thesimplest gauge invariant nonlocal condensate

〈q(0)E(0, x;A)q(x)〉|(xA)=0 = 〈q(0)q(x)〉

=∞∑n=0

1n!xµ1 . . . xµn〈qDµ1 . . . Dµnq〉 = 〈qq〉+

18x2〈qD2q〉+ . . . . (105)

The ratio

λ2 ≡ 〈qD2q〉

〈qq〉 (106)

can be interpreted as the average virtuality of the vacuum quarks. It is possible,however, to represent the 〈qD2q〉 operator in a different form:

〈qD2q〉 ≡ 〈qDµDµq〉 = 〈qDµDνq〉gµν = 〈q /D /Dq〉 − 〈qDµDνσµνq〉

= −m2q〈qq〉 −

12〈q [Dµ, Dν ]σµνq〉 = −m2

q〈qq〉+12〈qigGµνσµνq〉 (107)

where we used the identity

gµν = γµγν −γµγν − γνγµ

2= γµγν − σµν ,

the equation of motion/Dq(x) = −imqq(x) (108)

and the definition of the field strength tensor

[Dµ, Dν ] = −igGµν. (109)

Thus, the “average virtuality” of the vacuum quarks 〈qD2q〉 is directly relatedto the “average vacuum gluonic field strength” 〈qigGµνσµνq〉 . In many papersone can find the notation

〈qigGµνσµνq〉 ≡ m20〈qq〉. (110)

31

Using it, one can write the relation

λ2 =m2

0

2−m2

q. (111)

For light quarks, the m2q term can be neglected.

4.3 Equations of motion and the current conservation

The equation of motion /Dq(x) = −imqq(x) brings in correlations betweenmatrix elements of different operators. These correlations, in fact, are veryimportant to preserve the symmetry properties required by the structure ofthe currents j(x) in the correlators. For example, one might expect that acorrelator of the electromagnetic vector currents

Πµν(p) =∫〈T jµ(0)jν(x)〉eipx d4x (112)

should be transverse:

Πµν(p) = (pµpν − gµνp2) Π(p2) (113)

due to the current conservation condition ∂µjµ = 0. Indeed, for the pertur-bative one-loop contribution, using the explicit form of the propagator in theconfiguration space

S(x) ∝ −2i/x

x2

[1x2

+m2

4

]+m

x2+O(m3) (114)

one finds that the integrand is

1(x2)4

[gµνx

2 − 2xµxν]− m2

(x2)3

[gµνx

2 − 4xµxν]

+ . . . . (115)

It is easy to see that both terms satisfy the transversality condition

∂

∂xµ. . . = 0,

and this means that

Πpertµν (p) = (pµpν − gµνp2)Πpert(p2)

at least up to the third order in m. One can, in fact, perform the calculationsin the momentum space directly, to see that this property of Πpert

µν (p) is validto all orders in m.

32

Now, let us consider the simplest nonperturbative contribution. In themomentum representation it looks like

〈qγµ/p+m

p2 −m2γνq〉 ∼

mgµνp2〈qq〉 . (116)

However, it seems that there is no way to get the pµpν term, since there isonly one /p term in the numerator. To solve this puzzle, let us consider thiscontribution in the configuration representation. Up to O(m2)-terms one has

〈q(0)γµS(x)γνq(x)〉 ∝ 〈q(0)γµ

(−2i

/x

(x2)2+m

x2

)γνq(x)〉 +O(m2)

= − 2i(x2)2

〈q(0)γµ/xγνq(x)〉 +m

x2〈q(0)γµγνq(x)〉 +O(m2). (117)

Note, that vacuum averages of local operators with an odd number of indicesvanish, and only those with an even number survive. Incorporating the Lorentzinvariance, we obtain

〈q(0)γµγνq(0)〉 = gµν〈qq〉,〈q(0)γµγαγνq(0)〉 = 0,

〈q(0)γµγαγνDβq(0)〉 = (gµαgνσ + gναgµσ − gµνgασ)〈qγσDβq〉,〈qγσDβq〉 =

gσβ4〈 q /Dq〉. (118)

Finally, incorporating the equation of motion

〈q /Dq〉 = −im〈qq〉 (119)

one obtains the transverse result

〈q(0)γµS(x)γνq(x)〉 ∝ m

2(x2)2〈qq〉[gµνx2 + 2xµxν ]. (120)

It is instructive to analyze also the 〈GG〉 contribution. First, using theexponential representation, it is easy to see that for the closed loop the ex-ponentials corresponding to the upper and the lower lines cancel each other,and only the O(G) terms remain. This general result can be illustrated onthe example of one-loop diagrams with two vacuum gluons. To study, how thecancellation proceeds in this case, one should expand the exponential up toO(A2) terms:

S(x,A) = S(x)

1 + ig(xA) +(ig)2

2(xA)2 + . . .

,

S(−x,A) = S(−x)

1− ig(xA) +(ig)2

2(xA)2 + . . .

. (121)

33

Now, combining the contributions of the three relevant diagrams, one obtains

(ig)2

12

(xA)2 +12

(xA)2 − (xA)2

= 0, (122)

i.e., the A−dependence disappeared. To calculate the O(G)- terms, it is mostconvenient to use the Fock-Schwinger gauge (xA) = 0, in which

Aµ(x) = xν∫ 1

0

Gνµ(tx) tdt =∞∑n=0

xνxµ1 . . . xµn

(n+ 2)n!Gνµ;µ1...µn(0)

=12xνGνµ(0) + . . . . (123)

Using this representation, one can construct the Feynman rules for the verticescorresponding to Gµν , Gµν;µ1 , etc. For example, in the momentum represen-tation

xν2Gνµ ⇒

12iGνµ

∂

∂kν, (124)

and the insertion of such a vertex into the quark propagator gives

12i

/k

k2γµGνµ

∂

∂kν

/k

k2=i

2Gνµ

/k

k2γµ

/k

k2γν

/k

k2(125)

The result looks like an insertion of two A-fields, Aν and Aµ, antisymmetrizedwith respect to the interchange of µ and ν. There is an important thingto remember while constructing the diagrammatic technique for the G-typevertices:in the momentum representation they contain derivatives acting ontoall the propagators between the G-vertex and the Fock-Schwinger zero point.The G-inserted propagator looks simpler in the configuration space:

S(x,G) = − /x

2π2(x2)2− g

16π2x2xαεαβµνG

µνγβγ5 +O(z0)〈GG〉 + . . . (126)

where z0 is the Fock-Schwinger fixed point. If one takes z0 = 0, then theO(〈GG〉) term disappears. This means that in the gauge (xA) = 0 only thediagram with the G-insertions into different lines survives. Its contribution inthe configuration space is easily calculated by just multiplying the two S(x,G)propagators. Taking the trace over the γ-matrices, one obtains the result

− 2i3(16π2)2(x2)2

(gµνx2 + 2xµxν)〈g2G2〉 (127)

34

proportional to the same transverse structure as the contribution due to themq〈qq〉 operator. This is because both operators have the same dimension,equal to 4 in mass units.

The QCD sum rules for the “standard” mesons, like ρ, π,K, etc., are basedon the OPE including the operators of dimension 6, at least. One of theseoperators can be easily obtained from the Taylor expansion of the nonlocalquark condensate containing a γ-matrix

〈q(0)γσq(x)〉 = 〈qγσDαq〉xα +13!〈qγσDα1Dα2Dα3q〉xα1xα2xα3 + . . . . (128)

We recall that the terms, containing an even number of derivatives, vanishsince the total number of indices is odd in theses cases. Incorporating theLorentz invariance, we obtain

〈q(0)γσq(x)〉 =14xσ〈q /Dq〉+

1144

xσx2(〈q /DDαDαq〉

+〈q Dα /DDα〉+ 〈q DαDα /D〉). (129)

Next step is to use the translation invariance 〈∂µq(0) . . . q(0)〉 = 0 the equationof motion /Dq = −imqq and the commutation relation [Dα, Dµ] = igGαµ to get

〈q(0)γσq(x)〉 =−imxσ

4

(〈qq〉+ x2

12〈qD2q〉

)+ig

288xσx2〈qγµ[Dα, G

µα]q〉. (130)

The commutator [Dα, Gµα] is equal to the covariant derivative of the gluonic

field[Dα, G

µα] = Gµα ;α . (131)

The final step is to use the gluonic equation of motion

Gµα ;α = jµa τa ≡

∑i=u,d,s

〈qiγµτaqi〉τa (132)

analogous to the Maxwell equation. Thus, due to the equations of motion, theapparently two-quark operator produces a four-quark term:

〈q(0)γσq(x)〉 =−imxσ

4(〈qq〉+

x2

12〈q(0)D2q〉) +

ig

288xσx2〈qγµjµa τaq〉. (133)

5 QCD Sum Rules for ρ-meson

5.1 Dispersion relation

As pointed out by SVZ (whose analysis1 we closely follow in this section) weclosely to determine the basic parameters characterizing the ρ-meson, e.g., its

35

d)c)b)a)

qqqq

Figure 6: Perturbative contribution up to two loops.

mass mρ and its e+e− coupling constant gρ, one should consider the correlatorof two vector currents

j(ρ)µ =

12

(uγµu− dγµd) (134)

with the ρ-meson quantum numbers:

Π(ρ)µν (q) = (qµqν − q2gµν)Π(ρ)(q2) = i

∫eiqx〈0|T (j(ρ)

µ (x) j(ρ)ν (0) )| 0〉 d4x.

(135)Via the dispersion relation

Π(ρ)(q2 = −Q2) = Π(ρ)(0)− Q2

12π2

∫ ∞0

RI=1(s)s(s+Q2)

ds (136)

this correlator is related to the total cross section of the e+e−-annihilation tohadrons with the isospin I = 1, measured in units of the e+e− → µ+µ− crosssection :

RI=1 =σ(e+e− → hadrons, I = 1)

σ(e+e− → µ+µ−). (137)

Indeed, the j(ρ)µ coincides with the isovector part of the electromagnetic cur-

rent and is responsible, therefore, for the production of the hadrons with I = 1in the e+e−-annihilation. The correlator Π(ρ)

µν (q) contains, of course, onlythe hadronic part of the relevant factors, and that explains the division byσ(e+e− → µ+µ−) .

5.2 Operator product expansion

On the other hand, one can construct the operator product expansion for thecorrelator Π(ρ)

µν (q) calculating the relevant Feynman diagrams. To begin with,

36

31h)g)f)

q

q

q

e)

qq

c)

qqq

b) d)a)

Figure 7: Nonperturbative contributions up to dimension six.

one should calculate the perturbative diagrams shown in Fig.6 to extract theperturbative spectral density

ρpert(s) =32θ(s)

1 +

αsπ

(138)

which is essentially the isovector (I = 1 isospin projection) quark chargesquared (1

2 )2 multiplied by 2 (u- and d- quarks) and multiplied by the numberof colors (3). The simplest nonperturbative contribution is due to the quarkcondensate (see Fig.7a). It is proportional to the current quark masses:

Πqq ∼ 〈0|muuu+mddd|0〉(Q2)2

,

with mu ∼ 4MeV,md ∼ 7MeV . Hence, this contribution is rather smallnumerically. The contribution

ΠGG ∼ g2 〈0|GG|0〉(Q2)2

due to the gluon condensate is given by three diagrams shown in Fig.7b-d. Asmentioned at the end of the previous section, only the diagram 7b contributesin the Fock-Schwinger gauge xµA

µ(x) = 0. It is rather straightforward tocalculate the diagrams 7e,f proportional to

g2〈qqqq〉(Q2)3

.

37

However, the contributions of the same four-quark type can be obtained alsofrom the diagrams 7g,h which originally are bilinear in quark fields. Notice,however, the numbers 1 and 3 attached to the relevant lines going into thevacuum. They indicate how many covariant derivatives are acting on therelevant “vacuum” field. Thus, the matrix elements are 〈qγµDνGαβq〉 and〈qγµDν1Dν2Dν3q〉. As discussed at the end of Section 4, after vacuum aver-aging and commutations only the 〈qγµDνGµνq〉 component survives and, byequations of motion it is reduced to 〈0|qγµτajaµq|0〉 with

jaµ =∑

i=u,d,s

〈0|qiγµτaqi|0〉.

5.3 Basic QCD sum rule for the ρ-meson channel

After the Borel transformation that modifies the dispersion integral∫ρ(s)dss+Q2

→∫e−s/M

2ρ(s)

ds

M2(139)

and the power corrections

1(Q2)n

→ 1(n− 1)!(M2)n

(140)

one obtains the following QCD sum rule1:∫ ∞0

e−s/M2RI=1(s)ds =

32M2

[1 +

αs(M)π

+4π2

M4(〈muuu〉+ 〈mddd〉) +

+π2

3M4〈αsπGG〉 − 2π3

M6〈αs(uγαγ5τ

au− dγαγ5τad)2〉 −

− 4π3

9M6〈αs(uγατau+ dγατ

ad)∑

i=u,d,s

〈0|qiγµτaqi|0〉].(141)

It is the starting point for further analysis.

5.4 Fixing the condensates

Next step is to specify the numerical values of the matrix elements on the r.h.s.of this sum rule (same values will be used in all other sum rules).• The first matrix element is in fact fixed by the current algebra:

〈muuu〉+ 〈mddd〉 = −12f2πm

2π = −1.7 · 10−4GeV4 (142)

38

• The value of the gluonic condensate was extracted from the analysis of theQCD charmonium sum rules1

〈αsπGG〉 = 1.2 · 10−2GeV4. (143)

• The four-quark operators are first approximated by the value dictated by thevacuum dominance (or factorization) hypothesis:

〈ψAψBψCψD〉 = 〈ψAψB〉〈ψCψD〉 − 〈ψAψD〉〈ψCψB〉(144)

where the subscripts A,B,C,D include spin, color and flavour. Using theorthogonality relation

〈ψAψB〉 =δABN〈ψψ〉 (145)

one can reduce then any four quark operator to the 〈qq〉2 form. In our case

αs〈(uγαγ5τau− dγαγ5τ

ad)2〉|vacuumdominance =

=329αs〈qq〉2 ≈ 6.5 · 10−4GeV6 (146)

αs〈(dγατad)∑

i=u,d,s

〈qiγατaqi〉|vacuumdominance =

= −329αs〈qq〉2 ≈ −6.5 · 10−4GeV6, (147)

with the numerical values following from the estimate

αs〈qq〉2 ≈ −1.8 · 10−4GeV6 (148)

which can be justified using the above relation between the quark condensate,quark masses (mu +md = 11MeV ), experimental values of the pion mass andpion decay constant fπ, under the assumption that the relevant normaliza-tion point corresponds to αs = 0.7. In fact, this estimate amounts to fixingone of the most important input parameters of the QCD sum rule methodand the quoted value is supported by successful predictions for many differentresonances.

5.5 Qualitative analysis of the QCD sum rule

Using the vacuum dominance hypothesis, we can rewrite the sum rule in amore compact form:∫ ∞

0

e−s/M2RI=1(s)ds =

32M2

[1 +

αs(M)π

− 2π2f2πm

2π

M4+

39

+π2

3M4〈αsπGG〉 − 448π3

M6αs〈qq〉2

]. (149)

Substituting the numerical values of the condensate parameters gives∫ ∞0

e−s/M2RI=1(s)ds =

32M2

[1+

αs(M)π

+0.1(0.6GeV2

M2)2−0.14(

0.6GeV2

M2)3

].

(150)Our experience with the sum rules for the quantum-mechanical oscillator (seeSection 2) suggests that to calculate the ρ-meson mass we should incorporatealso the “daughter” sum rule resulting from the original one after the differen-tiation with respect to 1/M2:∫ ∞

0

e−s/M2sRI=1(s)ds =

32M4

[1+

αs(M)π−0.1(

0.6GeV2

M2)2+0.28(

0.6GeV2

M2)3

].

(151)Note now that power corrections in the sum rules above are small for the valuesof the Borel parameter M as low as M2 = m2

ρ = 0.6 GeV2:∫ ∞0

e−s/m2ρRI=1(s)ds =

32m2ρ[1 + 0.1 + 0.1− 0.14], (152)∫ ∞

0

e−s/m2ρsRI=1(s)ds =

32m4ρ[1 + 0.1− 0.1 + 0.28]. (153)

At such a low value of M2 one should expect that the integral over the physicalcross section is dominated by the lowest resonance, i.e., by the ρ-meson. Itscontribution to RI=1(s), in the limit of the vanishing width, can be written as

RI=1ρ (s) =

12π2m2ρ

g2ρ

δ(s−m2ρ). (154)

These observations allow one to make a rough estimate of the predictionsfollowing from the QCD sum rule. To this end we, following SVZ1

• neglect, for M2 = m2ρ, the power corrections in the r.h.s. of the sum rules,

• neglect, M2 = m2ρ, the higher state contributions in the l.h.s. This gives the

following relation12π2m2

ρ

g2ρ

e−1 ≈ 32m2ρ. (155)

Simplifying it results in a famous SVZ1 expression for the gρ constant in termsof fundamental constants π and e:

g2ρ

4π∼=

2πe≈ 2.3, (156)

to be compared with the experimental value 2.36± 0.18.

40

5.6 Fitting the spectrum parameters

The standard procedure is to extract the parameters of the hadronic spectrumrequiring the best agreement between the two sides of the sum rule. Of course,having only a few first terms of the operator product expansion one shouldnot hope to reproduce all the details of the resonance structure in a particularchannel. Only the gross features of the spectrum can be extracted from theQCD sum rules. So, let us take the ansatz “first narrow resonance plus con-tinuum” we discussed in the oscillator case. In fact, one should expect that inQCD this ansatz should work better, since the higher hadronic resonances arevery broad, and the total cross section of their production is known to be wellrepoduced by the parton model (i.e., perturbative) calculation. Specifically,we take

RI=1ρ (s) =

12π2m2ρ

g2ρ

δ(s−m2ρ) +

32

(1 +

αsπ

)θ(s− s0). (157)

The parameters to fit are the coupling constant g2ρ, the ρ-meson mass m2

ρ andthe effective continuum threshold s0.

The strategy will be just the same as used for the oscillator states:• transfer the continuum contribution to the right hand side,• write the original sum rule in the form

f2ρe−m2

ρ/M2

=32M2

(1− e−s0/M2

)+ “power corrections”, (158)

where, for brevity, we introduced the notation f2ρ =

12π2m2ρ

g2ρ

,• write the daughter sum rule in the form

f2ρm

2ρe−m2

ρ/M2

=32M4

(1− (1 +

s0

M2)e−s0/M

2)

+“power corrections”, (159)

• divide the second sum rule by the first to get the expression for m2ρ as a

function of s0 and M2 (for fixed values of the condensates),• look for the value of the effective threshold s0 that produces the most stablecurve for m2

ρ (recall that with all power corrections taken into account, andthe exact ansatz for the higher states, one should obtain an M2-independentresult for m2

ρ).In this way one would obtainb s0 ≈ 1.5 GeV2 and m2

ρ ≈ 0.58 GeV2. Nextstep is to extract g2

ρ from the first sum rule, using these values of s0 and m2ρ.

bWith all the numbers explicitly given, the readers are encouraged to perform the fittingas an excercise.

41

The value s0 ≈ 1.5 GeV2 provides the most stable curve for g2ρ, corresponding

to g2ρ ≈ 2.17. The result is rather sensitive to the m2

ρ-value. If one takes theexperimental value m2

ρ ≈ 0.6 GeV2, then the most stable curve still is that fors0 ≈ 1.5GeV2, but the g2

ρ-value is a little higher g2ρ ≈ 2.40 and closer to the

experimental one g2ρ ≈ 2.36± 0.18. Anyway, it should be emphasized that all

the g2ρ-values extracted from the QCD sum rule differ from the experimental

value by less than 10%, the claimed precision of the method.

6 QCD Sum Rules for Pion

6.1 Axial vector current

Vector currents are a special case, because the relevant correlators can be re-lated to directly measurable cross sections. From purely theoretical point ofview, other currents such as scalar, pseudoscalar, axial vector etc., are notworse, but experimental information in these cases is normally limited to themasses of the lowest states. Still, the axial vector current is in a better situa-tion. First, its projection onto the pion state is just the pion decay constantfπ:

〈0|dγ5γµu|P 〉 = ifπPµ. (160)

Second, measuring the decays of the heavy lepton τ → ντ+X , one can measurethe coupling of a hadronic state X to the axial current. Our goal now is toshow how one can use the QCD sum rules to study the hadronic spectrumin the axial vector channel. Again, our presentation here closely follows theclassic SVZ analysis1.

Since the axial vector current aµ(x) = dγ5γµu is not conserved, the relevantcorrelator Π(ax)

µν (q) is a sum of two independent functions:

Π(ax)µν (q) = i

∫eiqx〈0|T (a+

µ (x) a−ν (0) )| 0〉 d4x = −gµνΠ1(Q2) + qµqνΠ2(Q2),

(161)or, in terms of the transverse and longitudinal components

Π(ax)µν (q) = −

gµν −

qµqνq2Π1(Q2)− qµqν

q2Π1(Q2) +Q2Π2(Q2)

. (162)

6.2 Massless pion

The common belief is that the pion mass vanishes in the limit of exact chiralsymmetry so that the pion is a Goldstone particle. It can be shown, thatthe operator product expansion in this case really requires the presence of a

42

massless particle in the limit mq → 0. In this limit the imaginary part ofthe longitudinal function Π1(Q2) + Q2Π2(Q2) ≡ Π|| vanishes, and Π|| is apolynomial in Q2. Now, let us switch on a small quark mass and keep termslinear in this mass. The statement is that the function Π1(Q2) +Q2Π2(Q2) isexactly calculable in this approximation:

Π1(Q2) +Q2Π2(Q2) =(mu +md)〈uu+ dd〉

Q2+O(m2

q). (163)

There are no terms linear in mq of higher orders in 1/Q2. Let us show thatthis equation really implies the existence of a nearly massless pion. First weshould express the Π-functions in terms of the hadronic contributions, usingthe dispersion relations:

Π1(Q2) =1π

∫ ∞0

Im Π1

s+Q2ds+ “subtr.” (164)

Π2(Q2) =1π

∫ ∞0

Im Π2

s+Q2ds+ “subtr.” (165)

Π1(Q2) +Q2Π2(Q2) =1π

∫ ∞0

Im Π1 − sIm Π2

s+Q2ds+ “subtr.” (166)

where Im Π1 contains contributions only from pseudovector (spin-1) states andtheir contribution should be transverse, i.e., proportional to (gµν − qµqν/q2):

Im Π1 =∑A

πf2Aδ(s−M2), (167)

while Im Π2 contains contributions both from pseudovector (spin-1) states andfrom the pseudoscalar (spin-0) ones and their contribution into Π(ax)

µν (q) shouldbe longitudinal i.e., proportional to qµqν :

sImΠ2 =∑A

πf2Aδ(s−M2

A) +∑P

πf2Pm

2P δ(s−M2

P ). (168)

Hence, only pseudoscalar states contribute to the longitudinal component ofΠ(ax)µν . For small quark masses we have

(mu +md)〈uu+ dd〉Q2

+O(m2q) = −

∑P

f2Pm

2P

Q2 +m2P

=

= −∑P

(f2Pm

2P

Q2− f2

Pm4P

Q4+ . . .

). (169)

43

All O(Q−4) terms should be O(m2q). This means that

f2Pm

4P = O(m2

q) (170)

for all states. Hence, the particles with masses that remain constant asmq → 0,decouple in the chiral limit:

m2P = O(m0

q), (171)

f2P = O(m2

q). (172)

The last relation also means that they do not contribute to the term linear inmq. The only way to get this term is to assume that there exists a state forwhich the O(mq)-term is brought by the m2

P -factor:

m2P = O(m1

q), (173)

fP = O(m0q). (174)

Such a state is naturally identified with the pion. In addition, we get thewell-known relation

f2πm

2π = −(mu +md)〈uu+ dd〉 (175)

demonstrating explicitly the vanishing of the pion mass in the chiral limit.To derive the starting statement, one should consider qµqνΠ(ax)

µν (q). Byequations of motion it is related to the correlator of the pseudoscalar densities:

qµqνΠ(ax)µν (q) = Q2(Π1(Q2) +Q2Π2(Q2)) =

= −i∫eiqx〈0|T

((u(x)γ5d(x)) (d(0)γ5u(0))

)| 0〉 d4x+ const. (176)

The constant on the right-hand side is due to contact terms which normallyarise if one differentiates a T -product. This constant corresponds to the 1/Q2-term in the combination Π1(Q2) +Q2Π2(Q2). Direct calculation of this termgives the result stated in the beginning of this subsection.

Thus, the operator product expansion implies that m2π → 0 in the mq → 0

limit, and , hence, the pion mass should not be fitted from the sum rule.Within our accuracy one can set m2

π = 0 from the start.

6.3 QCD sum rule for the axial channel

The operator product expansion for the correlator of two axial currents is com-pletely analogous to that of two vector currents, the only difference being the

44

presence of the extra γ5-matrices in the vertices corresponding to the currents.For massless quarks, however, this produces no changes both in perturbativeand O(〈GG〉) contributions. However, the four-quark terms shown in Fig.7e,fchange sign, because one of the γ-matrix type terms is substituted by thecondensate factor 〈qq〉 which has the structure of 1 with respect to the Diracindices.

After the Borel transformation the sum rule looks as follows:∫ ∞0

e−s/M2Im Π2(s)ds =

M2

4π

[1 +

αs(M)π

+π2

3M4〈αsπGG〉+

+4π3

M6〈αs(uγαγ5τ

ad dγαγ5τau)〉+

+4π3

9M6〈αs(uγατau+ dγατ

ad)∑

q=u,d,s

〈0|qγµτaq|0〉]. (177)

In the axial channel we have two states that are the lowest among otherstates with the same quantum numbers: a pseudoscalar (the pion) and anaxial vector (A1-meson) ones. Thus, one can try the ansatz with two low-lyingstates (“pion+A1+continuum”):

Im Π2(s) = πf2πδ(s) + πf2

A1δ(s−m2

A1) +

14π

(1 +

αsπ

)θ(s > s

(A1)0 ). (178)

One can also treat the A1 as a part of the continuum. This, more crudeapproximation, corresponds to the simplified ansatz

Im Π2(s) = πf2πδ(s) +

14π

(1 +

αsπ

)θ(s > s

(π)0 ). (179)

The constants fπ, fA1 are defined as follows:

〈0|dγ5γµu|π〉 = ifπPµ, 〈0|dγ5γµu|A1〉 = ifA1εµ, (180)

where εµ is the A1 polarization vector.

6.4 Fitting the sum rule

Substituting the numerical values of the condensates into the basic sum rulegives ∫ ∞

0

e−s/M2Im Π2(s)ds =

M2

4πr[1 +

αs(M)π

+0.1(0.6GeV2

M2)2 + 0.22(

0.6GeV2

M2)3

]. (181)

45

The “daughter” sum rule is

∫ ∞0

e−s/M2Im Π2(s)sds =

M2

4π

[1 +

αs(M)π

−0.1(0.6GeV2

M2)2 − 0.44(

0.6GeV2

M2)3

]. (182)

Note, that in the m2π = 0 approximation the pion does not contribute to the

daughter sum rule, and the spectrum structure for sIm Π2(s) = Im Π1(s) ismore like in the vector channel: first non-zero mass resonance + continuum.This is reflected in the structure of the relevant sum rule: power correctionsare negative. They are considerably larger than those in the ρ-channel, andone should expect that the A1-mass (squared) is much larger than m2

ρ.

The fitting procedurec is most straightforward in the case of the simplifiedansatz. One should transfer the continuum contribution to the right hand sideto get f2

π as a function of s(π)0 and the Borel parameter M2:

4π2f2π = M2

(1− e−s

(π)0 /M2

) [1 +

αs(M)π

+π2

3M2〈αsπGG〉+

704π3

M4αs〈qq〉2

].

(183)The most stable curve corresponds to s0 ≈ 0.7 GeV2 and 4π2f2

π ≈ 0.65 GeV2,the experimental value being 4π2f2

π = 0.67 GeV2.

One can also try the ansatz in which A1 is treated as a (narrow) resonance.The strategy is just the same as in the ρ-case. To this end, one should• differentiate the daughter (second) sum rule with respect to 1/M2 to get thethird sum rule,• transfer the continuum contribution to the right hand side in all sum rules,• extract m2

A1and s

(A1)0 from the ratio of the third sum rule to the second,

• using the values m2A1≈ 1.6 GeV2 and s

(A1)0 ≈ 2.4 GeV2 obtained from the

previous fitting, find f2A1

from the second sum rule,• using the value 4π2f2

A1≈ 1.95 GeV2 obtained from the previous fitting, find

f2π from the first sum rule. The value f2

π ≈ 0.69 GeV2 obtained in this wayis very close to the experimental one f2

π |exp = 0.67 GeV2. The A1 mass valuem2A1≈ 1.6 GeV2 extracted from the QCD sum rule also is in a very good

agreement with experiment.

cAgain, we encourage the readers to obtain the curves themselves.

46

7 QCD Sum Rules for Pion Form Factor

7.1 Three-point function

To analyze the pion form factor, we consider the three-point function, i.e., thecorrelator of the three currents20,21:

T µαβ(p1, p2) = i

∫e−ip1x+ip2y〈jβ(y)Jµ(0)j+

α (x)〉d4xd4y, (184)

whereJµ =

23uγµu− 1

3dγµd (185)

is the electromagnetic current and

jα = dγ5γαu (186)

is the axial current we used in the preceding section to study the static prop-erties of the pion. The pion contribution

〈0|jβ(y)|p2〉〈p2|Jµ(0)|p1〉〈p1|j+α (x)|0〉 (187)

into this three-point correlator can be obtained by inserting the pion state(s)between the currents. The form factor Fπ(q2) appears in the middle matrixelement:

〈p2|Jµ(0)|p1〉 = (pµ1 + pµ2 )Fπ(q2), (188)

while the other two are projections of the axial current onto the initial andfinal pion states

〈0|jβ(0)|p2〉 = ifπ(p2)β (189)〈p1|j+

α (0)|0〉 = ifπ(p1)α. (190)

The three-point function T µαβ(p1, p2) is the sum of different structures eachcharacterized by the relevant invariant amplitude ti(p2

1, p22, q

2). The first idea isto study the pion form factor analizing the invariant amplitude correspondingto the structure (p2)β(p1)α(pµ1 + pµ2 ). However, there are other structures((p1)β(p2)α, (p1)β(p1)α, (p2)β(p2)α that coincide with (p2)β(p1)α in the limitq → 0. This complication disappears if all the basic structures are expandedin P = (p1 + p2)/2 and q = p2 − p1. Then the pion contribution is

(p2)β(p1)α(pµ1 + pµ2 ) = 2(Pβ +

qβ2

)(Pα −

qα2

)Pµ = 2PβPαPµ +O(q),

47

and the “best” structure (not changing when q varies) is PαPβPµ. The simplestway to extract the relevant amplitude (hereafter referred to as T ) is to multiplyT µαβ(p1, p2) by nµn

αnβ , where n is a lightlike vector orthogonal to q. Theproperty n2 = 0 kills all the structures containing gαβ and other g-terms,while the property (nq) = 0 kills all the structures containing any q-factor.Thus, the amplitude we are going to analyze is

T (p21, p

22, q

2) =nµn

αnβ

2(nP )3T µαβ(p1, p2). (191)

7.2 Dispersion representation

To extract information about the form factors of physical states, one shouldincorporate the double dispersion relation

T (p21, p

22, q

2) =1π2

∫ ∞0

ds1

∫ ∞0

ds2ρ(s1, s2, q

2)(s1 − p2

1)(s2 − p22)

+ “subtractions”.

(192)The subtraction terms, in general, are polynomial in p2

1 and/or p22. They dis-

appear after one applies B(p21 → M2

1 )B(p22 → M2

2 ), the Borel transformationin both variablesd. The double Borel transform Φ = B1B2T is related to thedouble spectral density ρ(s1, s2, q

2 = −Q2) by

Φ(M21 ,M

22 , q

2 = −Q2) =1π2

∫ ∞0

ds1

M21

∫ ∞0

ds2

M22

exp(− s1

M21

− s2

M22

)ρ(s1, s2, Q

2).

(193)

7.3 Operator product expansion

In contrast to the two-point correlators Π(Q2) analyzed in the previous Section,the three-point amplitudes T (p2

1, p22, q

2) have three arguments. To incorporateasymptotic freedom, one should take the invariants p2

1, p22 corresponding to the

“hadronized” channels to be negative and sufficiently large, say, |p21|, |p2

2| >1 GeV2. Depending on the value of the third invariant q2 that works as anexternal parameter (it also should be negative, of course), there are threeessentially different situations.• The simplest is the symmetric case whenQ2 is of the same order of magnitudeas |p2

1|, |p22|:

|p21| ∼ |p2

2| ∼ Q2 ∼ µ2.

dIn our specific case, the subtractions are exhausted by a constant, and a single differen-tiation with respect to any variable is sufficient to remove it.

48

d)c)b)

+ +

q

2pp

1

+=

a)

Figure 8: Structure of OPE in case of symmetric kinematics.

The operator product expansion for T in this case has the following structure:

T (p21, p

22, q

2) =∑i

1(µ2)di

Ci

(µ2

p21

,µ2

p22

,µ2

q2

)〈Oi〉 (194)

where Ci/(µ2)di is a perturbatively calculable short-distance coefficient func-tion, 〈Oi〉 is the vacuum matrix element of a local operator , i.e., some con-densate term and di is the dimension of Oi in mass units (see Fig.8).• The case when Q2 is small is more complicated:

|p21| ∼ |p2

2| ∼ µ2, Q2 µ2.

The coefficient functions Ci calculated in the symmetric kinematics, mightbecome singular in this limit because of (µ2/Q2)n- or log(µ2/Q2)-terms. Theoperator product expansion in such a situation has a modified form:

T (p21, p

22, q

2) =∑i

1(µ2)di

Ci

(µ2

p21

,µ2

p22

,µ2

q2

)〈Oi〉

+∑j

1(µ2)dj

Ek(µ2

p21

,µ2

p22

)Πk(Q2) (195)

where the coefficient functions Ci are regular in the Q2 → 0 limit, Πk(Q2) isthe correlator of the electromagnetic current J and a local operator Ok

Πk(Q2) = i

∫e−iqx〈Ok(x)J(0)〉 d4x,

dk is the dimension of Ok in mass units. The coefficient functions Ek are (theFourier transforms of) those for the operator product expansion of two axialcurrents

T (j j+) =∑k

EkOk.

49

In the modified OPE for the three-point function, the contributions of thefirst type correspond to situation when large momentum flow connects all threeexternal verteces, while those of the second type describe the situation whenlarge momentum flows only between the vertices x and y related to the axialcurrents. The momenta in the J-channel are of the order of Q, i.e., small, andthe correlator Πj(Q2) is a nonperturbative object accumulating long-distanceinformation.• In the opposite limit of large Q2

|p21| ∼ |p2

2| ∼ µ2, Q2 µ2,

a simple expansion in powers of 1/|p21|, 1/|p2

2 (i.e.,in 1/µ2) might be destabi-lized by large (Q2/µ2)n-terms. Appearance of terms growing with Q2 is a bitsurprizing, since form factors should decrease as Q2 →∞. But the appearanceof such terms is just the artifact of the 1/µ2 expansion: even a very reasonablefunction like 1/(Q2 + µ2), vanishing as Q2 → ∞, produces a divergent serieswhen expanded in 1/µ2. Thus, a possible way out in the large-Q2 limit is aresummation of the parametrically enhanced contributions of (Q2/µ2)n-type.In fact, it can be established, that such contributions result from the Taylorexpansion of the original nonlocal combinations 〈ϕ(0)ϕ(x)〉 (nonlocal conden-sates), and the idea is to construct sum rules without expanding the nonlocalcondensates, treating them as some phenomenological functions describing thepropagation of quark and gluons in the physical vacuum.

7.4 Perturbative contribution

The structure of the power corrections in the operator product expansion, as wediscussed above, depends on the particular Q2-region under study. However,the starting point of any expansion is the perturbative contribution, in ourcase corresponding to the triangle diagram plus radiative corrections to it.

Using Feynman parametrization, one can rather easily obtain the followingrepresentation21 for the double Borel transform of the lowest-order perturbativetriangle diagram (see Fig.8a)

Φpert(M21 ,M

22 , Q

2) =3

2π2(M21 +M2

2 )

∫ 1

0

x(1− x)

exp− xQ2

(1− x)(M21 +M2

2 )

I(x,M2

1 ,M22 ,m

2q, eq) dx (196)

where I(x,M21 ,M

22 ,m

2q, eq) is the factor containing the dependence on the

50

quark masses

I(x,M21 ,M

22 ,m

2q, eq) = eu exp

−(m2u

1− x +m2d

x

)(1M2

1

+1M2

2

)+ u↔ d.

(197)In view of extreme smallness of the u- and d-quark masses, this factor can besafely set to unity. The x-variable may be interpreted as the fraction of thetotal momentum P carried by the passive quark (active is the quark interactingwith the EM current). This integral representation is very convenient to studythe behavior of Φpert(M2

1 ,M22 , Q

2) for small and large Q2. In the low-Q2 limitwe get

Φpert(M21 ,M

22 , Q

2) =1

4π2(M21 +M2

2 )

1− 2

Q2

M21 +M2

2

(198)

−∞∑N=0

( −Q2

M21 +M2

2

)N+2N + 3N !

(log[

Q2

M21 +M2

2

]+

1N + 3

− ψ(N + 1))

.

The first two terms of the expansion can be obtained by simply expanding theexponential factor in Q2 under the integral sign. However, the third term ofsuch an expansion produces a logarithmically divergent integral

. . .

∫ 1

0

12!

(Q2

M21 +M2

2

)2x3

1− x dx,

and higher terms have power divergences when integrated in the region x→ 1.This region corresponds to the situation when the whole large momentum Pis carried by the passive quark, and the active quark carries small momen-tum proportional to q. This is precisely the situation producing additionalterms in the operator product expansion at small Q2. The logarithmic factorlog(Q2), singular when Q2 → 0, just signalizes that there are long-distanceeffects (quarks propagate over distances ∼ 1/Q) limiting the applicability ofperturbation theory.

In the large-Q2 limit one should expand Φpert(M21 ,M

22 , Q

2) in 1/Q2 ratherthan in Q2:

Φpert(M21 ,M

22 , Q

2) =3

2π2

M21 +M2

2

(Q2)2

1− 8

M21 +M2

2

Q2+

+16

∞∑N=2

(−)N (N + 1)(N + 3)!(M2

1 +M22

Q2

)N. (199)

51

Thus, the perturbative contribution vanishes like 1/Q4 asQ2 →∞. Combiningthe spectral representation for Φ(M2

1 ,M22 , Q

2) and the explicit expression forit, we get the equation for the perturbative spectral density ρpert(s1, s2, Q

2)

1π2

∫ ∞0

ds1

M21

∫ ∞0

ds2

M22

exp(− s1

M21

− s2

M22

)ρpert(s1, s2, Q2) =

=3

2π2(M21 +M2

2 )

∫ 1

0

x(1− x) exp(− xQ2

(1− x)(M21 +M2

2 )

)dx. (200)

Solving it (this amounts to a double inverse Laplace transformation) gives theperturbative spectral density:

ρpert(s1, s2, Q2) =

32

[(Q2)2

2!+

(Q2)3

3!

]1√

(s1 + s2 +Q2)2 − 4s1s2

. (201)

In the low-Q2 limit the perturbative spectral density can be expanded in Q2:

ρpert(s1, s2, Q2)|Q2→0 =

14δ(s1 − s2)θ(s1)θ(s2) +

+Q2

4(s1 + s2)δ′′(s1 − s2)θ(s1)θ(s2) + . . . . (202)

For large Q2 one can expand it in 1/Q2:

ρpert(s1, s2, Q2)|Q2→∞ =

32θ(s1) θ(s2)

s1 + s2

Q4− 4

s21 + s2

2 + 4s1s2

Q6+ . . .

.

(203)Thus, for small Q2, the perturbative spectral density is concentrated on theline s1 = s2 : there are no transitions between states with different masses.Along this line, the spectral density does not vary: it has no dependence ons1 + s2. In the opposite limit, when Q2 → ∞, the situation is, in a sense,reversed: the density is nonzero in the whole region s1 > 0, s2 > 0, thereis no dependence on (s1 − s2), and the density grows like s1 + s2. The lastobservation means that higher states are relatively more important for largeQ2 than for the small ones.

7.5 Sum rule for the pion form factor at moderate Q2

The physical spectral density ρ(s1, s2, Q2), of course, differs from its pertur-

bative analog. It has a rather complicated structure on the s1, s2-plane. Inparticular, ρ(s1, s2, Q

2) contains the term corresponding to the pion form fac-tor

ρππ(s1, s2, Q2) = π2f2

πFπ(Q2)δ(s1 −m2π)δ(s2 −m2

π). (204)

52

In addition, it contains the contributions corresponding to transitions betweenthe pion and higher resonances, and also the terms related to elastic and tran-sition form factors of the higher resonances. Constructing the two-dimensional(s1, s2) analog of the simplest ansatz “lowest state plus continuum” we cantreat all the contributions, except for the ρππ, as “continuum”, i.e., assumethat ρ(s1, s2, Q

2) = ρpert(s1, s2, Q2) outside the square (0, s0)× (0, s0):

ρ(s1, s2, Q2) = ρππ(s1, s2, Q

2) + (1− θ(s1 < s0)θ(s2 < s0)) ρpert(s1, s2, Q2).

(205)Now, transferring the continuum contribution to the right-hand side of the sumrule and using the operator product expansion in the symmetric kinematics(Q2 ∼M2

1 ,M22 ), we obtain the QCD sum rule for the pion form factor:

f2πFπ(Q2) =

1π2

∫ s0

0

ds1

∫ s0

0

ds2 exp(−s1 + s2

M2

)ρpert(s1, s2, Q

2) +

+αs〈GG〉12πM2

+1681παs〈qq〉2M4

(13 + 2

Q2

M2

). (206)

To treat initial and final states on equal footing and to simplify the analysis,we took M2

1 = M22 . Unfortunately, the double integral in this sum rule cannot

be calculated explicitly for arbitrary Q2. However, it is instructive to considerthe formal limit Q2 = 0:

f2πFπ(0) =

14π2

M2

2

(1− e−2s0/M

2)

+αs〈GG〉12πM2

+20881

παs〈qq〉2M4

. (207)

This should be compared to the f2π sum rule:

f2π =

m2

4π2

(1− e−s0/m2

)+αs〈GG〉12πm2

+17681

παs〈qq〉2m4

. (208)

The condensate terms look almost identical, but comparing the perturbativeterms we observe that M2, the Borel parameter of the form factor sum rule,should be larger by factor 2 than m2, the Borel parameter of the f2

π sum rule:M2 = 2m2. After this change, the 〈GG〉-term in the form factor sum rule ismultiplied by 1

2 and the 〈qq〉2-term is multiplied by 14

f2πFπ(0) =

m2

4π2

(1− e−s0/m2

)+αs〈GG〉24πm2

+5281παs〈qq〉2m4

. (209)

Thus, we obtained a sum rule, which looks like that for f2π , but with the gluonic

condensate term smaller by factor 2 ≈ (1.4)2, and the quark condensate term

53

smaller by factor ∼ 3.5 ≈ (1.5)3. Since it is the values of the condensates that(after the fitting procedure) determine all the hadronic scales, we conclude thatall the hadronic paramerters having the dimension of (mass)2, i.e., s0 and thecombination f2

πFπ(0), extracted from this sum rule are by a factor 1.4 − 1.5smaller than sπ0 and f2

π, respectivelye. To get s0 closer to sπ0 , one should restorethe matching between the perturbative and the condensate terms, i.e., decreasethe perturbative term by a factor 2 to 4. This allows us to make an educatedguess that QCD sum rule for the pion form factor will most effectively workin the intermediate region where the form factor varies between 0.5 and 0.25,i.e., for Q2 between 0.5 GeV2 and 2 GeV2.

The readers can check these guesses by explicit fitting procedure. To thisend, one should plot the combination f2

πFπ(Q2) as a function of the Borel pa-rameter M2 for different values of the effective threshold s0. As a “true” valueof s0 one can take the value for which the curve is constant as M2 →∞. Onecan observe that the value of s0 obtained in this way slowly grows with Q2,from 0.6 GeV2 for Q2 = 0.5 GeV2 to 1.0 GeV2 for Q2 = 3 GeV2. The growth ofs0 reflects the fact that, when the Q2 value increases, the condensate contribu-tions remain constant (with the quark term even slowly growing) whereas theperturbative term decreases. The constancy of the condensate contributions isan artifact of our approximation. If one includes operators of higher dimensionsin the OPE, there appear terms diminishing the total condensate contributionfor large Q2. These terms, however, have the structure (Q2/M2)n, and oneshould resum them to get a reasonable result. However, if one restricts theanalysis to the region 0.5 GeV2 < Q2 < 3 GeV2, the predictions of our sumrule agree with existing pion form factor data within 10%− 20% accuracy.

7.6 Local quark-hadron duality for the pion form factor

The approach described above allows to obtain the form factor at differentvalues of Q2 by independent fitting procedure, point by point. But it istempting to see the Q2-curve as a whole. This can be done by looking atthe Fπ(Q2)-curves at fixed s0 for different values of the Borel parameter M2.For M2 > 1.5 GeV2, all the curves are close to each other, approaching the lim-iting (s0-dependent) form as M2 →∞. The nature of this convergence is verysimple: the condensate terms disappear from the sum rule in the M2 → ∞

eOne should not worry about the fact that a straightforward extrapolation of the formfactor sum rule to the point Q2 = 0, gives Fπ(0) ≈ 0.7. This sum rule is valid only forsufficiently high Q2, and, as we emphasized, there appear additional terms in the OPEwhich will increase the condensate contributions to make them equal to those in the f2

π sumrule and recover the Fπ(0) = 1 normalization condition.

54

limit, and the limiting curve corresponds to the local duality relation:

f2πFπ(Q2) =

1π2

∫ s0

0

ds1

∫ s0

0

ρpert(s1, s2, Q2) ds2. (210)

Note, that the function ρpert(s1, s2, Q2) describes the transition from a free-

quark (qq) state with invariant mass s1 to a similar state with mass s2. Thelocal duality relation states, essentially, that one can calculate the pion formfactor by averaging the form factors of such transitions over the appropriateduality interval s0. In the two-point case, the local duality produces the rela-tion between the pion decay constant f2

π and the duality interval s0

s0 = 4π2f2π ≈ 0.7 GeV2. (211)

We recall that the duality interval is the effective threshold for production ofhigher resonances, and is very natural that it is just in the middle betweenm2π ≈ 0 and m2

A1≈ 1.6 GeV2. Using the explicit form for ρpert(s1, s2, Q

2), weobtain21

Fπ(Q2)|local duality = 1− 1 + 6s0/Q2

(1 + 4s0/Q2)3/2. (212)

This formula gives a correct value for the form factor at Q2 = 0, but thepredicted slope is wrong:

Fπ(Q2)|local duality = 1− 34Q√s0

+ . . . . (213)

This is just the same situation we encountered applying the local duality tothe form factor of the lowest oscillator state. The reason is that there is aserious mismatch between perturbative and physical spectral densities: theperturbative density is concentrated in the narrow stripe |s1 − s2| < Q nearthe s1 = s2 line, while physical density, in addition to the elastic form factors(points at the s1 = s2 line), contains also transition form factors (line s1 =0, s2 > s(3π threshold), etc.) corresponding to the regions very far from the s1 =s2 line. In other words, our model for the hadronic spectrum is too rough atsmall nonzeroQ2. The perturbative density in this region can be approximatedby

ρpert(s1, s2, Q2) ∼ 1

Qθ(|s1 − s2| < Q). (214)

Integrating it over any line s1 + s2 =const., one obtains a result analytic inQ2. In our model, however, there are small triangles not included into theintegration region. Their area is ∼ Q2, and the density is ∼ 1/Q, so the

55

resulting contribution (to be subtracted from 1) is O(Q) = O(√Q2). To avoid

this, one can try a “triangle” model for the higher states: “pion + (continuumoutside the triangle formed by the line (s1 + s2 = S0)”. The local dualityrelation in this case gives21

Fπ(Q2)triangle l.d. =S0

8π2f2π(1 +Q2/2S0)2

. (215)

To get the correct normalization Fπ(0) = 1 one should take S0 = 8π2f2π = 2s0.

However, the slope in this case is too small F tr.l.d.π (Q2)|S0=2s0 = 1−Q2/2s0, byfactor 2 smaller than the experimental one, and the curve goes well above boththe experimental points and the curve corresponding to the “squared” form ofthe local duality. To match with the latter, one should take S0 =

√2s0. In

this case the area of the local duality region is the same. Outside the small-Q2

region, there is good agreement between the two local duality curves and thedata approximated by the fit F fitπ (Q2) ≈ 1/(1 +Q2/0.47).

7.7 One-gluon exchange contribution

The contribution into the pion form factor we calculated using the lowest orderO(α0

s) term in the perturbative spectral density corresponds to the so-called“soft diagram”. Since this contribution decreases as 1/Q4 for large Q2, it isnormally ignored in the perturbative QCD analysis concentrated on the studyof the “hard gluon exchange diagram” that has the 1/Q2 asymptotic behaviordictated by the quark counting rules.

However, the soft contribution is very close to the experimental data leav-ing not much place for any other contribution. To estimate the contributiondue to the “hard gluon” exchange diagrams, one should include the O(αs(Q2)-terms in the perturbative spectral density. To simplify the analysis, one canuse the local duality approximation. Still, the two-loop calculation is rathercomplicated, but the results, with a rather high accuracy, can be approximatedby a simple interpolation

1π2f2

π

∫ s0

0

ds1

∫ s0

0

ds2∆ρpert(s1, s2, q2) ≈ αs

π

11 +Q2/2s0

(216)

between the Q2 = 0 value (related by the Ward identity to the O(αs) term ofthe 2-point correlator) and the asymptotic behavior F asπ (Q2) = 8παs(Q2)f2

π/Q2.

56

pp

qq

b)a)

Figure 9: Gauge-invariant set of 〈qq〉2 diagrams.

7.8 Pion form factor at small Q2

As we discussed earlier, if one simple-mindedly extrapolates the QCD sum rule,derived in the moderate-Q2 region, to the point Q2, one obtains the expression

f2πFπ(0) ?=

m2

4π2

(1− e−s0/m2

)+αs〈GG〉24πm2

+5281παs〈qq〉2m4

(217)

that differs from the sum rule for f2π

f2π =

m2

4π2

(1− e−s0/m2

)+αs〈GG〉12πm2

+17681

παs〈qq〉2m4

. (218)

The condensate terms have the coefficients essentially smaller than it is nec-essary to enjoy the simple Ward identity prediction Fπ(Q2) = 1. The latterrequires that the sum rule for f2

πFπ(0) should coincide with that for f2π . The

Ward identity, being just the statement that the total electric charge of thepion is equal to that of the electron, should not be violated. Hence thereshould exist the additional condensate contributions “visible” at Q2 = 0 (tosatisfy the Ward identity) and dying rapidly for Q2 > m2

ρ (to reproduce theintermediate-Q2 sum rule).

To get a feeling about how these terms might look like, consider the αs〈qq〉2diagrams. To enjoy the consequences of the Ward identity, one should have agauge invariant set of diagrams. This can be achieved by inserting the photonvertex into all possible lines, including also the “external” lines correspondingto quarks going into vacuum. However, in perturbation theory, this bringsin terms ∼ 1/Q2 singular in the Q2 → 0 limit. Such a “coefficient function”cannot be calculated perturbatively. Thus, one should separate the large- andsmall-momentum parts of the diagram

T (p2, q2) = C(p2)⊗Π(q2) (219)

57

and treat the Q2-dependent part as a correlator of the original electromagneticcurrent J and some local operator constructed from the quark fields going “intovacuum” (see Fig.9b):

Π(q2) ∼∫

eiqx〈T (Jµ(0)O(x))〉d4x. (220)

The question now is whether it is possible to calculate contributions like thecorrelators Π(q2) at small q? There are essentially two types of these correlatorsdepending on the type of the operator O. First, one should take into accountthat under the T -product sign, one cannot throw out the operators apparentlyvanishing due to the equations of motion, e.g., those containing /D acting onthe quark field, since there appear contact terms. In this case Π(q2) is aconstant proportional to the quark condensate 〈qq〉. In all other cases one can(approximately) calculate Π(q2) using the following QCD sum rule strategy:• Calculate Π(q2) for large Q2 using the operator product expansion. Thisnormally gives

Π(q2 = −Q2) ∼ 〈qq〉Q2

+ k〈qD2q〉Q4

+ . . . .

• Use the dispersion relation

Π(Q2) =1π

∫ ∞0

ρ(s) dss+Q2

.

• Assume an appropriate ansatz for the spectral density:

ρ(s) = πAδ(s−m2ρ) + πBδ(s −m2

R).

Since the perturbative density is zero in this case, the higher states (lyingabove the ρ-meson) are approximated by an effective resonance R.• Requiring the best agreement between the two representations for Π(Q2),extract the values of the constants A,B and m2

R.• The result is

Π(Q2) =A

Q2 +m2ρ

+B

Q2 +m2R

,

and one can use it for small Q2.In this way one can obtain the QCD sum rule for the pion form factor

valid in the low-Q2 region22:

f2πFπ(Q2) = “perturbative part” +

αs〈GG〉12πm2

[12

+m2ρ

2(m2ρ +Q2)

]+

58

+1681παs〈qq〉2m4

[4 + 6

(1.6m2

ρ

m2ρ +Q2

− 0.6m2R

m2R +Q2

)+

m2ρ

m2ρ +Q2

+Q2

m2

]. (221)

This sum rule possesses all the required properties:• For Q2 = 0 its left hand side exactly coincides with that of the f2

π sum rule.• additional terms die out when Q2 gets large.• The curve for the pion form factor calculated with the help of this sum rulein the region of small Q2, at the matching point Q2 = m2

ρ, agrees within 10%with the curve obtained from the moderate-Q2 sum rule.• the pion charge radius, extracted from the low-Q2 sum rule,

〈r2π〉1/2|SR = 0.66± 0.03 fm

is in good agreement with the experimental value

〈r2π〉1/2|exp. = 0.636± 0.036 fm .

Thus, the QCD sum rules give an accurate description of the pion formfactor in the whole region where the reliable data exist. They allow one totrace its behavior from the normalization point Q2 = 0, through the low-Q2

region, to the region of moderately large Q2. Everywhere the QCD sum ruleresults are in agreement with experiment within the accuracy of the method.The QCD sum rules unambigously demonstrate that, in the experimentallyaccessible region, the form factor is dominated by the soft contribution , notcalculable within a purely perturbative approach. They show that the one-gluon-exchange mechanism, though dominant in the asymptotic limit, is ofminor importance in the region Q2 < 4 GeV2 and, maybe, till even largervalues of Q2.

Acknowledgements

This work was supported by the US Department of Energy under contractDE-AC05-84ER40150.

References

1. M.A. Shifman, A.I. Vainshtein and V.I. Zakharov: Nucl. Phys. B 147,385, 448 (1979).

2. A.V. Efremov and A.V. Radyushkin, Phys. Lett. B 94, 245 (1980);Theor. Mat. Fiz. 42, 147 (1980).

59

3. G.P. Lepage and S.J. Brodsky, Phys. Lett. B 87, 359 (1979);Phys. Rev. D 22, 2157 (1980).

4. V.L. Chernyak, A.R. Zhitnitsky and V.G. Serbo, JETP Lett. 26, 594(1977); V.L. Chernyak and A.R. Zhitnitsky: JETP Lett. 25, 510 (1977);Yad. Fiz. 31, 1053 (1980); Phys. Rep. 112, 173 (1984).

5. G.R. Farrar and D.R. Jackson, Phys. Rev. Lett. 43, 246 (1979).6. G.P. Lepage and S.J. Brodsky, Phys. Lett. B 87, 359 (1979)7. A.I. Vainshtein, V.I. Zakharov, V.A. Novikov, M.A. Shifman,

Sov.J.Nucl.Phys. 32, 840 (1980).8. V.A. Beilin, A.V. Radyushkin, Form-factors of oscillator states from

QCD sum rules, report JINR-P2-87-216, Dubna (1987).9. A.V. Radyushkin, Habilitation dissertation, JINR, Dubna (1987), un-

published.10. V.A. Nesterenko and A.V. Radyushkin, Sov.J.Nucl.Phys. 39, 811 (1984).11. A.V. Radyushkin, Phys. Lett. B 271, 218 (1991).12. K. Wilson, Phys. Rev. 179, 1499 (1969).13. A.V. Efremov and A.V. Radyushkin, Theor. Mat. Fiz. 44, 774 (1981).14. S. N. Nikolaev and A.V. Radyushkin, Phys.Lett. 110B, 476 (1982);

Nucl. Phys. B 213, 285 (1983).15. V.A. Fock, Sov. Phys., 12 (1937).16. J. Schwinger, Phys. Rev. 82, 664 (1951).17. M.S. Dubovikov and A.V. Smilga, Nucl.Phys. B 185, 109 (1981).18. M.A. Shifman, Nucl. Phys. B 173, 13 (1980).19. S.V. Ivanov, G.P. Korchemsky and A.V. Radyushkin,

Sov.J.Nucl.Phys. 44, 145 (1986).20. B.L. Ioffe and A.V. Smilga, Phys.Lett. B 114, 53 (1982);

Nucl. Phys. B 216, 373 (1983).21. V.A. Nesterenko and A.V. Radyushkin, Phys.Lett. B 115, 410 (1982).22. V.A. Nesterenko and A.V. Radyushkin, JETP Lett. 39, 707 (1984).

60